From 1928904316e80ba0549b94ae1f905d7e79bda812 Mon Sep 17 00:00:00 2001 From: Michele Di Giorgio Date: Wed, 3 Feb 2021 16:05:00 +0000 Subject: Make NEON Pooling kernels and functions state-less Partially resolves COMPMID-3999 Change-Id: Ib39d40694df5c5f0a9401488e0c3af3ac26e8c55 Signed-off-by: Michele Di Giorgio Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4984 Tested-by: Arm Jenkins Reviewed-by: Georgios Pinitas Comments-Addressed: Arm Jenkins --- Android.bp | 7 +- .../runtime/NEON/functions/NEPoolingLayer.h | 18 +- docs/00_introduction.dox | 6 +- src/core/NEON/NEKernels.h | 1 - src/core/NEON/kernels/NEFillBorderKernel.cpp | 26 +- src/core/NEON/kernels/NEFillBorderKernel.h | 14 +- src/core/NEON/kernels/NEMaxUnpoolingLayerKernel.h | 4 +- src/core/NEON/kernels/NEPoolingLayerKernel.cpp | 2612 -------------------- src/core/NEON/kernels/NEPoolingLayerKernel.h | 229 -- .../assembly/NEPoolingAssemblyWrapperKernel.cpp | 269 -- .../assembly/NEPoolingAssemblyWrapperKernel.h | 116 - .../kernels/CpuPoolingAssemblyWrapperKernel.cpp | 276 +++ .../cpu/kernels/CpuPoolingAssemblyWrapperKernel.h | 123 + src/core/cpu/kernels/CpuPoolingKernel.cpp | 2605 +++++++++++++++++++ src/core/cpu/kernels/CpuPoolingKernel.h | 226 ++ .../NEON/functions/NEPoolingAssemblyDispatch.cpp | 156 -- .../NEON/functions/NEPoolingAssemblyDispatch.h | 149 -- src/runtime/NEON/functions/NEPoolingLayer.cpp | 103 +- src/runtime/cpu/operators/CpuPooling.cpp | 130 + src/runtime/cpu/operators/CpuPooling.h | 102 + .../cpu/operators/CpuPoolingAssemblyDispatch.cpp | 98 + .../cpu/operators/CpuPoolingAssemblyDispatch.h | 89 + 22 files changed, 3717 insertions(+), 3642 deletions(-) delete mode 100644 src/core/NEON/kernels/NEPoolingLayerKernel.cpp delete mode 100644 src/core/NEON/kernels/NEPoolingLayerKernel.h delete mode 100644 src/core/NEON/kernels/assembly/NEPoolingAssemblyWrapperKernel.cpp delete mode 100644 src/core/NEON/kernels/assembly/NEPoolingAssemblyWrapperKernel.h create mode 100644 src/core/cpu/kernels/CpuPoolingAssemblyWrapperKernel.cpp create mode 100644 src/core/cpu/kernels/CpuPoolingAssemblyWrapperKernel.h create mode 100644 src/core/cpu/kernels/CpuPoolingKernel.cpp create mode 100644 src/core/cpu/kernels/CpuPoolingKernel.h delete mode 100644 src/runtime/NEON/functions/NEPoolingAssemblyDispatch.cpp delete mode 100644 src/runtime/NEON/functions/NEPoolingAssemblyDispatch.h create mode 100644 src/runtime/cpu/operators/CpuPooling.cpp create mode 100644 src/runtime/cpu/operators/CpuPooling.h create mode 100644 src/runtime/cpu/operators/CpuPoolingAssemblyDispatch.cpp create mode 100644 src/runtime/cpu/operators/CpuPoolingAssemblyDispatch.h diff --git a/Android.bp b/Android.bp index bc5ae34bba..5653fc8a63 100644 --- a/Android.bp +++ b/Android.bp @@ -289,7 +289,6 @@ cc_library_static { "src/core/NEON/kernels/NENormalizationLayerKernel.cpp", "src/core/NEON/kernels/NEPadLayerKernel.cpp", "src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp", - "src/core/NEON/kernels/NEPoolingLayerKernel.cpp", "src/core/NEON/kernels/NEPriorBoxLayerKernel.cpp", "src/core/NEON/kernels/NEQLSTMLayerNormalizationKernel.cpp", "src/core/NEON/kernels/NEQuantizationLayerKernel.cpp", @@ -340,7 +339,6 @@ cc_library_static { "src/core/NEON/kernels/arm_gemm/quantized.cpp", "src/core/NEON/kernels/arm_gemm/rowsum_indirect_s8.cpp", "src/core/NEON/kernels/arm_gemm/rowsum_indirect_u8.cpp", - "src/core/NEON/kernels/assembly/NEPoolingAssemblyWrapperKernel.cpp", "src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp", "src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp", "src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp", @@ -409,6 +407,8 @@ cc_library_static { "src/core/cpu/kernels/CpuFillKernel.cpp", "src/core/cpu/kernels/CpuFloorKernel.cpp", "src/core/cpu/kernels/CpuPermuteKernel.cpp", + "src/core/cpu/kernels/CpuPoolingAssemblyWrapperKernel.cpp", + "src/core/cpu/kernels/CpuPoolingKernel.cpp", "src/core/cpu/kernels/CpuReshapeKernel.cpp", "src/core/cpu/kernels/CpuSubKernel.cpp", "src/core/cpu/kernels/activation/NEON/fp16.cpp", @@ -736,7 +736,6 @@ cc_library_static { "src/runtime/NEON/functions/NEPermute.cpp", "src/runtime/NEON/functions/NEPhase.cpp", "src/runtime/NEON/functions/NEPixelWiseMultiplication.cpp", - "src/runtime/NEON/functions/NEPoolingAssemblyDispatch.cpp", "src/runtime/NEON/functions/NEPoolingLayer.cpp", "src/runtime/NEON/functions/NEPriorBoxLayer.cpp", "src/runtime/NEON/functions/NEQLSTMLayer.cpp", @@ -796,6 +795,8 @@ cc_library_static { "src/runtime/cpu/operators/CpuFill.cpp", "src/runtime/cpu/operators/CpuFloor.cpp", "src/runtime/cpu/operators/CpuPermute.cpp", + "src/runtime/cpu/operators/CpuPooling.cpp", + "src/runtime/cpu/operators/CpuPoolingAssemblyDispatch.cpp", "src/runtime/cpu/operators/CpuReshape.cpp", "src/runtime/cpu/operators/CpuSub.cpp", "src/runtime/gpu/cl/operators/ClActivation.cpp", diff --git a/arm_compute/runtime/NEON/functions/NEPoolingLayer.h b/arm_compute/runtime/NEON/functions/NEPoolingLayer.h index d239138165..91b3a709f4 100644 --- a/arm_compute/runtime/NEON/functions/NEPoolingLayer.h +++ b/arm_compute/runtime/NEON/functions/NEPoolingLayer.h @@ -32,17 +32,15 @@ namespace arm_compute { +// Forward declarations class ITensor; class ITensorInfo; -class NEPoolingLayerKernel; -class NEFillBorderKernel; -class NEPoolingAssemblyDispatch; /** Basic function to simulate a pooling layer with the specified pooling operation. This function calls the following NEON kernels: * * -# @ref NEFillBorderKernel (executed if padding size is different from zero) - * -# @ref NEPoolingLayerKernel - * -# @ref NEPoolingAssemblyDispatch + * -# @ref cpu::kernels::CpuPoolingKernel + * -# @ref cpu::CpuPoolingAssemblyDispatch */ class NEPoolingLayer : public IFunction { @@ -86,14 +84,8 @@ public: void run() override; private: - std::shared_ptr _memory_manager; - - std::unique_ptr _pooling_layer_kernel; - std::unique_ptr _border_handler; - std::unique_ptr _asm_glue; - - bool _is_global_pooling_layer; - DataLayout _data_layout; + struct Impl; + std::unique_ptr _impl; }; } #endif /* ARM_COMPUTE_NEPOOLINGLAYER_H */ diff --git a/docs/00_introduction.dox b/docs/00_introduction.dox index 735f60ad2c..ab2495dbff 100644 --- a/docs/00_introduction.dox +++ b/docs/00_introduction.dox @@ -166,7 +166,7 @@ v20.11 Public major release - NELocallyConnectedMatrixMultiplyKernel - @ref NEGEMMLowpOffsetContributionKernel - @ref NEGEMMTranspose1xWKernel - - @ref NEPoolingLayerKernel + - NEPoolingLayerKernel - @ref NEConvolutionKernel - @ref NEDepthwiseConvolutionLayerNativeKernel - @ref NEGEMMLowpMatrixMultiplyKernel @@ -1120,7 +1120,7 @@ v18.01 Public maintenance release - Added QASYMM8 support to the following NEON kernels: - NEDepthwiseConvolutionLayer3x3Kernel - @ref NEFillBorderKernel - - @ref NEPoolingLayerKernel + - NEPoolingLayerKernel - Added new examples: - graph_cl_mobilenet_qasymm8.cpp - graph_inception_v3.cpp @@ -1299,7 +1299,7 @@ v17.03 Sources preview - New NEON kernels / functions: - NEActivationLayerKernel / @ref NEActivationLayer - GEMM refactoring + FP16 support (Requires armv8.2 CPU): @ref NEGEMMInterleave4x4Kernel, @ref NEGEMMTranspose1xWKernel, @ref NEGEMMMatrixMultiplyKernel, @ref NEGEMMMatrixAdditionKernel / @ref NEGEMM - - @ref NEPoolingLayerKernel / @ref NEPoolingLayer + - NEPoolingLayerKernel / @ref NEPoolingLayer v17.02.1 Sources preview - New OpenCL kernels / functions: diff --git a/src/core/NEON/NEKernels.h b/src/core/NEON/NEKernels.h index 87eec38605..c636e5b3be 100644 --- a/src/core/NEON/NEKernels.h +++ b/src/core/NEON/NEKernels.h @@ -101,7 +101,6 @@ #include "src/core/NEON/kernels/NENormalizationLayerKernel.h" #include "src/core/NEON/kernels/NEPadLayerKernel.h" #include "src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.h" -#include "src/core/NEON/kernels/NEPoolingLayerKernel.h" #include "src/core/NEON/kernels/NEPriorBoxLayerKernel.h" #include "src/core/NEON/kernels/NEQLSTMLayerNormalizationKernel.h" #include "src/core/NEON/kernels/NEQuantizationLayerKernel.h" diff --git a/src/core/NEON/kernels/NEFillBorderKernel.cpp b/src/core/NEON/kernels/NEFillBorderKernel.cpp index 488079062b..10384d4176 100644 --- a/src/core/NEON/kernels/NEFillBorderKernel.cpp +++ b/src/core/NEON/kernels/NEFillBorderKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2020 Arm Limited. + * Copyright (c) 2016-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -33,12 +33,8 @@ #include "src/core/NEON/kernels/NEFillBorderKernel.h" #include "src/core/helpers/WindowHelpers.h" -#include -#include - namespace arm_compute { -class Coordinates; namespace { inline void fill_constant_value_single_channel_special(ITensor *tensor, const Window &window, unsigned int right, unsigned int bottom, const PixelValue &constant_border_value) @@ -98,22 +94,28 @@ NEFillBorderKernel::NEFillBorderKernel() } void NEFillBorderKernel::configure(ITensor *tensor, BorderSize border_size, BorderMode border_mode, const PixelValue &constant_border_value) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(tensor); + _tensor = tensor; + configure(tensor->info(), border_size, border_mode, constant_border_value); +} + +void NEFillBorderKernel::configure(ITensorInfo *tensor, BorderSize border_size, BorderMode border_mode, const PixelValue &constant_border_value) { ARM_COMPUTE_ERROR_ON_NULLPTR(tensor); //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use NEON FP16 instructions. - ARM_COMPUTE_ERROR_ON(tensor->info()->data_type() == DataType::UNKNOWN); + ARM_COMPUTE_ERROR_ON(tensor->data_type() == DataType::UNKNOWN); - _tensor = tensor; _border_size = border_size; _mode = border_mode; _constant_border_value = constant_border_value; - _border_size.limit(tensor->info()->padding()); + _border_size.limit(tensor->padding()); Window win; win.set(Window::DimX, Window::Dimension(0, 1, 1)); win.set(Window::DimY, Window::Dimension(0, 1, 1)); - win.use_tensor_dimensions(_tensor->info()->tensor_shape(), Window::DimZ); + win.use_tensor_dimensions(tensor->tensor_shape(), Window::DimZ); INEKernel::configure(win); } @@ -156,6 +158,12 @@ void NEFillBorderKernel::run(const Window &window, const ThreadInfo &info) } } +void NEFillBorderKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) +{ + _tensor = tensors.get_tensor(TensorType::ACL_SRC_DST); + run(window, info); +} + void NEFillBorderKernel::fill_replicate_single_channel(const Window &window) { uint8_t *const start_valid_region = _tensor->ptr_to_element(_tensor->info()->valid_region().anchor); diff --git a/src/core/NEON/kernels/NEFillBorderKernel.h b/src/core/NEON/kernels/NEFillBorderKernel.h index 65908bebee..2c851583ed 100644 --- a/src/core/NEON/kernels/NEFillBorderKernel.h +++ b/src/core/NEON/kernels/NEFillBorderKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2020 Arm Limited. + * Copyright (c) 2016-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -65,9 +65,21 @@ public: * */ void configure(ITensor *tensor, BorderSize border_size, BorderMode border_mode, const PixelValue &constant_border_value = PixelValue()); + /** Initialise the function. + * + * @note This kernel fills the borders within the XY-planes. + * + * @param[in,out] tensor Tensor info to process. Data types supported: All. + * @param[in] border_size Size of the border to fill in elements. + * @param[in] border_mode Border mode to use for the convolution. + * @param[in] constant_border_value (Optional) Constant value to use for borders if border_mode is set to CONSTANT. + * + */ + void configure(ITensorInfo *tensor, BorderSize border_size, BorderMode border_mode, const PixelValue &constant_border_value = PixelValue()); // Inherited methods overridden: void run(const Window &window, const ThreadInfo &info) override; + void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override; private: void fill_replicate_single_channel(const Window &window); diff --git a/src/core/NEON/kernels/NEMaxUnpoolingLayerKernel.h b/src/core/NEON/kernels/NEMaxUnpoolingLayerKernel.h index 8cdfe2b953..f42272826c 100644 --- a/src/core/NEON/kernels/NEMaxUnpoolingLayerKernel.h +++ b/src/core/NEON/kernels/NEMaxUnpoolingLayerKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2020 Arm Limited. + * Copyright (c) 2020-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -56,7 +56,7 @@ public: * * @param[in] input Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. * @param[in] indices Tensor containing the offset to store the input elements in the output tensor. - * @ref NEPoolingLayerKernel with indices should precede this function in order to + * @ref cpu::kernels::CpuPoolingKernel with indices should precede this function in order to * properly reconstruct the output tensor. * The tensor shape of this tensor has to be equal to the input tensor shape. Data type supported: U32. * @param[out] output Destination tensor. Data types supported: Same as @p input. diff --git a/src/core/NEON/kernels/NEPoolingLayerKernel.cpp b/src/core/NEON/kernels/NEPoolingLayerKernel.cpp deleted file mode 100644 index b46843badd..0000000000 --- a/src/core/NEON/kernels/NEPoolingLayerKernel.cpp +++ /dev/null @@ -1,2612 +0,0 @@ -/* - * Copyright (c) 2017-2020 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/NEON/kernels/NEPoolingLayerKernel.h" - -#include "arm_compute/core/Error.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/ITensor.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Utils.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/core/Window.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "src/core/AccessWindowStatic.h" -#include "src/core/CPP/Validate.h" -#include "src/core/NEON/NEAsymm.h" -#include "src/core/NEON/NEFixedPoint.h" -#include "src/core/NEON/NEMath.h" -#include "src/core/helpers/AutoConfiguration.h" -#include "src/core/helpers/WindowHelpers.h" -#include "support/ToolchainSupport.h" - -#include "src/core/NEON/wrapper/wrapper.h" -#include -#include -#include -#include -#include -#include -#include - -namespace arm_compute -{ -using namespace misc::shape_calculator; - -namespace -{ -template -inline typename std::enable_if::value, int8_t>::type -quantize(float val, const UniformQuantizationInfo &info) -{ - return quantize_qasymm8_signed(val, info); -} - -template -inline typename std::enable_if::value, uint8_t>::type -quantize(float val, const UniformQuantizationInfo &info) -{ - return quantize_qasymm8(val, info); -} - -inline float calculate_avg_scale(bool exclude_padding, DataLayout data_layout, const Coordinates &id, const int pool_size_x, const int pool_size_y, const int upper_bound_w, const int upper_bound_h, - const int pad_x, const int pad_y, const int stride_x, const int stride_y) -{ - const unsigned int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - const unsigned int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); - - int start_x = id[idx_width] * stride_x - pad_x; - int start_y = id[idx_height] * stride_y - pad_y; - - const int end_x = std::min(start_x + pool_size_x, upper_bound_w); - const int end_y = std::min(start_y + pool_size_y, upper_bound_h); - if(exclude_padding) - { - start_x = std::max(0, start_x); - start_y = std::max(0, start_y); - } - return 1.f / ((end_y - start_y) * (end_x - start_x)); -} - -template -inline void scale_vector_q16x8(bool exclude_padding, TVec &v, const Coordinates &id, int id_offset, int step, - const int pool_size, const int upper_bound_w, const int upper_bound_h, - const int pad_x, const int pad_y, const int stride_x, const int stride_y) -{ - int start_x = (id.x() + id_offset) * stride_x - pad_x; - int start_y = id.y() * stride_y - pad_y; - const int end_y = std::min(start_y + pool_size, upper_bound_h); - if(exclude_padding) - { - start_y = std::max(0, start_y); - } - - std::array elems = - { - { - wrapper::vgetlane(v, 0), - wrapper::vgetlane(v, 1), - wrapper::vgetlane(v, 2), - wrapper::vgetlane(v, 3), - wrapper::vgetlane(v, 4), - wrapper::vgetlane(v, 5), - wrapper::vgetlane(v, 6), - wrapper::vgetlane(v, 7), - } - }; - - for(auto &el : elems) - { - int c_start_x = start_x; - const int end_x = std::min(c_start_x + pool_size, upper_bound_w); - if(exclude_padding) - { - c_start_x = std::max(0, c_start_x); - } - float scale = 1.f / ((end_y - start_y) * (end_x - c_start_x)); - el *= scale; - start_x += step * stride_x; - } - - v = wrapper::vsetlane(elems[0], v, 0); - v = wrapper::vsetlane(elems[1], v, 1); - v = wrapper::vsetlane(elems[2], v, 2); - v = wrapper::vsetlane(elems[3], v, 3); - v = wrapper::vsetlane(elems[4], v, 4); - v = wrapper::vsetlane(elems[5], v, 5); - v = wrapper::vsetlane(elems[6], v, 6); - v = wrapper::vsetlane(elems[7], v, 7); -} - -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, - unsigned int &pooled_w, unsigned int pooled_h, const ITensorInfo *indices, Size2D pool_size) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); - - int pool_stride_x = 0; - int pool_stride_y = 0; - PoolingType pool_type = pool_info.pool_type; - const PadStrideInfo pad_stride_info = pool_info.pad_stride_info; - std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); - - ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input); - if(indices) - { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(indices, 1, DataType::U32); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(pool_type != PoolingType::MAX, "Pooling indices only supported for MAX pooling method"); - } - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON(pool_type == PoolingType::L2 && is_data_type_quantized(input->data_type())); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_quantized(input->data_type()) && !pool_info.exclude_padding && (pool_info.pool_type == PoolingType::AVG) && pool_info.pad_stride_info.has_padding() - && (input->data_layout() == DataLayout::NHWC), - "exclude_padding equal false is not supported for AVG Pooling with padding on quantized types"); - - if(output->total_size() != 0) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output); - ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH)) != pooled_w) - || (output->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT)) != pooled_h)); - - if(indices) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG((pool_size != Size2D(2, 2)), "Pooling indices only supported for pool size 2x2"); - ARM_COMPUTE_RETURN_ERROR_ON((indices->dimension(get_data_layout_dimension_index(indices->data_layout(), DataLayoutDimension::WIDTH)) != pooled_w) - || (indices->dimension(get_data_layout_dimension_index(indices->data_layout(), DataLayoutDimension::HEIGHT)) != pooled_h)); - } - } - - return Status{}; -} - -Status validate_arguments_pool_info(const unsigned int pool_size_x, const unsigned int pool_size_y) -{ - ARM_COMPUTE_RETURN_ERROR_ON(pool_size_x == 0); - ARM_COMPUTE_RETURN_ERROR_ON(pool_size_y == 0); - - return Status{}; -} - -std::pair validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, ITensorInfo *indices, const PoolingLayerInfo &pool_info, - unsigned int &num_elems_processed_per_iteration, - BorderSize &border_size, - unsigned int pooled_w, unsigned int pooled_h, int pool_size_x, int pool_size_y) -{ - // Output auto inizialitation if not yet initialized - auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_pool_shape(*input, pool_info))); - if(indices) - { - // Indices auto inizialitation if not yet initialized - auto_init_if_empty(*indices, (input->clone()->set_tensor_shape(compute_pool_shape(*input, - pool_info))) - .set_data_type(DataType::U32) /* we store the offset to the element */); - } - const auto data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? input->data_layout() : pool_info.data_layout; - unsigned int num_elems_read_per_iteration = 0; - unsigned int num_elems_horizontal_window = 0; - int pool_stride_x = 0; - int pool_stride_y = 0; - const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); - const int input_width = input->dimension(idx_width); - const int input_height = input->dimension(idx_height); - const PadStrideInfo pad_stride_info = pool_info.pad_stride_info; - std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); - const int pool_pad_right = pad_stride_info.pad_right(); - const int pool_pad_top = pad_stride_info.pad_top(); - const int pool_pad_left = pad_stride_info.pad_left(); - const int pool_pad_bottom = pad_stride_info.pad_bottom(); - const bool is_square = pool_size_x == pool_size_y; - - // Check output dimensions - std::tie(pooled_w, pooled_h) = scaled_dimensions(input->dimension(idx_width), - input->dimension(idx_height), - pool_size_x, - pool_size_y, - pad_stride_info); - - //If it's not squared and optimized will be executed the MxN - num_elems_read_per_iteration = 1; - num_elems_processed_per_iteration = 1; - num_elems_horizontal_window = 1; - - if(is_square) - { - switch(input->data_type()) - { - case DataType::QASYMM8: - case DataType::QASYMM8_SIGNED: - switch(pool_size_x) - { - case 2: - num_elems_read_per_iteration = 16; - num_elems_processed_per_iteration = (pool_stride_x == 2) ? 8 : 15; - num_elems_horizontal_window = (pool_stride_x == 2) ? 8 : 16; - break; - case 3: - num_elems_read_per_iteration = 16; - num_elems_processed_per_iteration = (pool_stride_x == 2) ? 7 : 14; - num_elems_horizontal_window = (pool_stride_x == 2) ? 8 : 16; - break; - default: - break; - } - break; -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - case DataType::F16: - switch(pool_size_x) - { - case 2: - case 3: - num_elems_read_per_iteration = 4; - num_elems_processed_per_iteration = 1; - num_elems_horizontal_window = 1; - break; - default: - break; - } - break; -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - case DataType::F32: - switch(pool_size_x) - { - case 2: - num_elems_read_per_iteration = 2; - break; - case 3: - num_elems_read_per_iteration = 4; // We use vload4 for pooling3 - break; - case 7: - num_elems_read_per_iteration = 8; // We use vload8 for pooling7 - break; - default: - break; - } - num_elems_processed_per_iteration = 1; - num_elems_horizontal_window = 1; - break; - default: - ARM_COMPUTE_ERROR("Element size not supported"); - break; - } - } - - bool window_changed = false; - Window win{}; - if(data_layout == DataLayout::NCHW) - { - // Number of iterations in X dimension - const int num_iterations_x = (pooled_w + num_elems_processed_per_iteration - 1) / num_elems_processed_per_iteration; - // Upper limit for the number of right/bottom border elements that are accessed - const int upper_bound_w = ((num_iterations_x - 1) * num_elems_processed_per_iteration * pool_stride_x - pool_pad_left + num_elems_read_per_iteration) - input_width; - const int upper_bound_h = ((pooled_h - 1) * pool_stride_y - pool_pad_top + pool_size_y) - input_height; - border_size = BorderSize(pool_pad_top, pool_pad_right, pool_pad_bottom, pool_pad_left); - border_size.right = std::max(upper_bound_w, pool_pad_right); - border_size.bottom = std::max(upper_bound_h, pool_pad_bottom); - TensorShape output_shape{ input->tensor_shape() }; - output_shape.set(0, pooled_w); - output_shape.set(1, pooled_h); - TensorInfo output_info(input->clone()->set_tensor_shape(output_shape)); - win = calculate_max_window(output_info, Steps(num_elems_processed_per_iteration)); - AccessWindowStatic input_access(input, -pool_pad_left, -pool_pad_top, input_width + border_size.right, input_height + border_size.bottom); - AccessWindowHorizontal output_access(output, 0, num_elems_horizontal_window); - if(indices) - { - AccessWindowHorizontal indices_access(indices, 0, num_elems_horizontal_window); - window_changed = update_window_and_padding(win, input_access, output_access, indices_access); - } - else - { - window_changed = update_window_and_padding(win, input_access, output_access); - } - output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); - } - - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - return std::make_pair(err, win); -} - -template -inline T vcvtq_q32_f32(float32x4_t values); - -template <> -inline uint32x4_t vcvtq_q32_f32(float32x4_t values) -{ - return vcvtq_u32_f32(values); -} - -template <> -inline int32x4_t vcvtq_q32_f32(float32x4_t values) -{ - return vcvtq_s32_f32(values); -} - -template -inline float32x4_t vcvtq_f32_q32(T values); - -template <> -inline float32x4_t vcvtq_f32_q32(uint32x4_t values) -{ - return vcvtq_f32_u32(values); -} - -template <> -inline float32x4_t vcvtq_f32_q32(int32x4_t values) -{ - return vcvtq_f32_s32(values); -} - -template -inline Tout vrequantize_pooling_with_scale(const float32x4x4_t &acc, const float quant_rescale, const float scale_pooling, const int32_t new_offset); - -template <> -inline uint8x16_t vrequantize_pooling_with_scale(const float32x4x4_t &acc, const float quant_rescale, const float scale_pooling, const int32_t new_offset) -{ - const float new_scale = quant_rescale / scale_pooling; - return vquantize(acc, UniformQuantizationInfo(new_scale, new_offset)); -} - -template <> -inline int8x16_t vrequantize_pooling_with_scale(const float32x4x4_t &acc, const float quant_rescale, const float scale_pooling, const int32_t new_offset) -{ - const float new_scale = quant_rescale / scale_pooling; - return vquantize_signed(acc, UniformQuantizationInfo(new_scale, new_offset)); -} - -template -inline Tout vrequantize_pooling(Tin vec1, Tin vec2, const UniformQuantizationInfo &requant_qinfo); - -template <> -inline uint8x16_t vrequantize_pooling(uint8x8_t vec1, uint8x8_t vec2, const UniformQuantizationInfo &requant_qinfo) -{ - const float32x4x4_t acc = - { - { - vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8((vec1))))), - vcvtq_f32_u32(vmovl_u16(vget_high_u16(vmovl_u8((vec1))))), - vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8((vec2))))), - vcvtq_f32_u32(vmovl_u16(vget_high_u16(vmovl_u8((vec2))))), - } - }; - return vquantize(acc, requant_qinfo); -} - -template <> -inline int8x16_t vrequantize_pooling(int8x8_t vec1, int8x8_t vec2, const UniformQuantizationInfo &requant_qinfo) -{ - const float32x4x4_t acc = - { - { - vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8((vec1))))), - vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8((vec1))))), - vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8((vec2))))), - vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8((vec2))))), - } - }; - return vquantize_signed(acc, requant_qinfo); -} - -template -inline T vrequantize_pooling(T &vec, const UniformQuantizationInfo &requant_qinfo); - -template <> -inline uint8x8_t vrequantize_pooling(uint8x8_t &vec, const UniformQuantizationInfo &requant_qinfo) -{ - const float32x4x2_t acc = - { - { - vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8((vec))))), - vcvtq_f32_u32(vmovl_u16(vget_high_u16(vmovl_u8((vec))))), - } - }; - return vquantize(acc, requant_qinfo); -} - -template <> -inline int8x8_t vrequantize_pooling(int8x8_t &vec, const UniformQuantizationInfo &requant_qinfo) -{ - const float32x4x2_t acc = - { - { - vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8((vec))))), - vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8((vec))))), - } - }; - return vquantize_signed(acc, requant_qinfo); -} - -} // namespace - -NEPoolingLayerKernel::NEPoolingLayerKernel() - : _func(nullptr), _input(nullptr), _output(nullptr), _indices(nullptr), _pool_info(), _data_layout(DataLayout::UNKNOWN), _num_elems_processed_per_iteration(0), _border_size(0), _is_square(false) -{ -} - -BorderSize NEPoolingLayerKernel::border_size() const -{ - return _border_size; -} - -void NEPoolingLayerKernel::configure(const ITensor *input, ITensor *output, const PoolingLayerInfo &pool_info, ITensor *indices) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - const PadStrideInfo pad_stride_info = pool_info.pad_stride_info; - const bool is_global_pooling = pool_info.is_global_pooling; - const int pool_stride_x = pad_stride_info.stride().first; - - // Get data layout - const auto data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? input->info()->data_layout() : pool_info.data_layout; - const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); - - // Update pool size in case of global pooling - const Size2D pool_size( - is_global_pooling ? input->info()->dimension(idx_width) : pool_info.pool_size.width, - is_global_pooling ? input->info()->dimension(idx_height) : pool_info.pool_size.height); - - // Validate pool info before calling scaled_dimensions - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_pool_info(pool_size.x(), pool_size.y())); - - // Check output dimensions - unsigned int pooled_w; - unsigned int pooled_h; - std::tie(pooled_w, pooled_h) = scaled_dimensions(input->info()->dimension(idx_width), - input->info()->dimension(idx_height), - pool_size.x(), - pool_size.y(), - pad_stride_info); - - // Perform validation step - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), pool_info, pooled_w, pooled_h, (indices) ? indices->info() : nullptr, pool_size)); - - // Set instance variables - _input = input; - _output = output; - _indices = indices; - _pool_info = pool_info; - _data_layout = input->info()->data_layout(); - _is_square = (pool_size.x() == pool_size.y()); - - // Get data type - const DataType data_type = input->info()->data_type(); - const bool is_nchw = _data_layout == DataLayout::NCHW; - - if(data_type == DataType::QASYMM8) - { - if(!is_nchw) - { - _func = &NEPoolingLayerKernel::poolingMxN_q8_nhwc; - } - else - { - if(pool_size.x() == 2 && pool_stride_x < 3 && _is_square) - { - _func = &NEPoolingLayerKernel::pooling2_q8_nchw; - } - else if(pool_size.x() == 3 && pool_stride_x < 3 && _is_square) - { - _func = &NEPoolingLayerKernel::pooling3_q8_nchw; - } - else - { - _func = &NEPoolingLayerKernel::poolingMxN_q8_nchw; - } - } - } - else if(data_type == DataType::QASYMM8_SIGNED) - { - if(!is_nchw) - { - _func = &NEPoolingLayerKernel::poolingMxN_q8_nhwc; - } - else - { - if(pool_size.x() == 2 && pool_stride_x < 3 && _is_square) - { - _func = &NEPoolingLayerKernel::pooling2_q8_nchw; - } - else if(pool_size.x() == 3 && pool_stride_x < 3 && _is_square) - { - _func = &NEPoolingLayerKernel::pooling3_q8_nchw; - } - else - { - _func = &NEPoolingLayerKernel::poolingMxN_q8_nchw; - } - } - } - else if(data_type == DataType::F16) - { - if(!is_nchw) - { - _func = &NEPoolingLayerKernel::poolingMxN_f16_nhwc; - } - else - { - if(_is_square) - { - switch(pool_size.x()) - { - case 2: - { - _func = &NEPoolingLayerKernel::pooling2_f16_nchw; - } - break; - case 3: - { - _func = &NEPoolingLayerKernel::pooling3_f16_nchw; - } - break; - default: - { - _func = &NEPoolingLayerKernel::poolingMxN_f16_nchw; - break; - } - } - } - else - { - _func = &NEPoolingLayerKernel::poolingMxN_f16_nchw; - } - } - } - else if(data_type == DataType::F32) - { - if(!is_nchw) - { - _func = &NEPoolingLayerKernel::poolingMxN_f32_nhwc; - } - else - { - if(_is_square) - { - switch(pool_size.x()) - { - case 2: - { - _func = &NEPoolingLayerKernel::pooling2_f32_nchw; - break; - } - case 3: - { - _func = &NEPoolingLayerKernel::pooling3_f32_nchw; - break; - } - case 7: - { - _func = &NEPoolingLayerKernel::pooling7_f32_nchw; - break; - } - default: - { - _func = &NEPoolingLayerKernel::poolingMxN_f32_nchw; - break; - } - } - } - else - { - _func = &NEPoolingLayerKernel::poolingMxN_f32_nchw; - } - } - } - - if(!is_nchw) - { - // Configure kernel window - Window win = calculate_max_window(*output->info(), Steps()); - Coordinates coord; - coord.set_num_dimensions(output->info()->num_dimensions()); - output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape())); - INEKernel::configure(win); - } - else - { - // Configure kernel window - auto win_config = validate_and_configure_window(input->info(), output->info(), (indices) ? indices->info() : nullptr, - pool_info, _num_elems_processed_per_iteration, _border_size, pooled_w, pooled_h, pool_size.x(), pool_size.y()); - ARM_COMPUTE_ERROR_THROW_ON(win_config.first); - INEKernel::configure(win_config.second); - } -} - -template -inline uint32_t offset_no_padding(uint32_t padded_offset, const Coordinates &id, const ITensorInfo &info, int pool_stride_x, int pool_stride_y) -{ - const int pad_left = info.padding().left; - const int pad_right = info.padding().right; - const int pad_top = info.padding().top; - const int pad_bottom = info.padding().bottom; - const int in_stride_y = static_cast(info.strides_in_bytes().y()); - const int in_stride_w = static_cast(info.strides_in_bytes()[3]); - const int pad_horiz = pad_left + pad_right; - const int pad_vert = pad_top + pad_bottom; - - if(info.data_layout() == DataLayout::NCHW) - { - const uint32_t offset_base = padded_offset - - sizeof(T) * pad_horiz * id.y() * pool_stride_y /* subtract padding elems per row */ - - pad_top * sizeof(T) /* top padding */ - - sizeof(T) * pad_horiz * info.tensor_shape()[1] * id.z() - pad_vert * in_stride_y * id.z() /* for each Z plane there are height*pad_right padding elems */ - - in_stride_w * id[3]; - - return offset_base; - } - else - { - const uint32_t offset_base = padded_offset - - sizeof(T) * pad_horiz * id.y() * pool_stride_x // subtract padding elems per row - - pad_top * sizeof(T) // top padding - - sizeof(T) * pad_horiz * info.tensor_shape()[1] * id.z() * pool_stride_y // for each Z plane there are width*pad_right padding elems - - in_stride_w * id[3]; - - return offset_base; - } -} - -template -void NEPoolingLayerKernel::pooling2_q8_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - Iterator input(_input, window_input); - Iterator output(_output, window); - - /** NEON vector types */ - using q8x8_t = typename wrapper::traits::neon_vector::type; - using q8x16_t = typename wrapper::traits::neon_vector::type; - using q8x8x2_t = typename std::conditional::value, uint8x8x2_t, int8x8x2_t>::type; - using q16_t = typename wrapper::traits::promote_t; - using q16x4_t = typename wrapper::traits::neon_vector::type; - using q16x8_t = typename wrapper::traits::neon_vector::type; - using q16x8x2_t = typename wrapper::traits::neon_vector::type; - - constexpr int pool_size = 2; - int pool_stride_x = 0; - int pool_stride_y = 0; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - const T *const input_top_ptr = reinterpret_cast(_input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top)))); - const T *const input_bottom_ptr = reinterpret_cast(_input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1))); - - const int scale_step_x = (pool_stride_x == 1) ? 2 : 1; - - const UniformQuantizationInfo input_qinfo = _input->info()->quantization_info().uniform(); - const UniformQuantizationInfo output_qinfo = _output->info()->quantization_info().uniform(); - const bool have_different_qinfo = input_qinfo != output_qinfo; - - const float requant_scale = output_qinfo.scale / input_qinfo.scale; - const int32_t requant_offset = output_qinfo.offset - static_cast(static_cast(input_qinfo.offset) / requant_scale); - const UniformQuantizationInfo requant_qinfo = UniformQuantizationInfo(requant_scale, requant_offset); - - execute_window_loop(window, [&](const Coordinates & id) - { - const auto top_data = wrapper::vloadq(input_top_ptr + input.offset()); - const auto bottom_data = wrapper::vloadq(input_bottom_ptr + input.offset()); - q8x8_t lower_res = {}; - q8x8_t upper_res = {}; - - if(pooling_type != PoolingType::MAX) - { - const q16x8x2_t top_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(top_data)), wrapper::vmovl(wrapper::vgethigh(top_data)) } }; - const q16x8x2_t bottom_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(bottom_data)), wrapper::vmovl(wrapper::vgethigh(bottom_data)) } }; - - // Add rows - const q16x8x2_t vrsum = - { - { - wrapper::vadd(top_data_q16.val[0], bottom_data_q16.val[0]), - wrapper::vadd(top_data_q16.val[1], bottom_data_q16.val[1]), - } - }; - - // Pair-wise add row data - const q16x4_t vpsum_1 = wrapper::vpadd(wrapper::vgetlow(vrsum.val[0]), wrapper::vgethigh(vrsum.val[0])); - const q16x4_t vpsum_2 = wrapper::vpadd(wrapper::vgetlow(vrsum.val[1]), wrapper::vgethigh(vrsum.val[1])); - - q16x8_t res_lower = wrapper::vcombine(vpsum_1, vpsum_2); - - // Scale lower result - scale_vector_q16x8(exclude_padding, res_lower, id, 0, scale_step_x, - pool_size, upper_bound_w, upper_bound_h, - pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - lower_res = wrapper::vmovn(res_lower); - - // Compute upper result for stride_x == 1 - if(pool_stride_x == 1) - { - // Shifted row sum - const q16x8x2_t vrsum_shifted = - { - { - wrapper::vext_1(vrsum.val[0], vrsum.val[1]), - wrapper::vext_1(vrsum.val[1], vrsum.val[1]) - } - }; - - // Pair-wise add shifted row - q16x8_t res_upper = wrapper::vcombine( - wrapper::vpadd(wrapper::vgetlow(vrsum_shifted.val[0]), wrapper::vgethigh(vrsum_shifted.val[0])), - wrapper::vpadd(wrapper::vgetlow(vrsum_shifted.val[1]), wrapper::vgethigh(vrsum_shifted.val[1]))); - - // Scale upper result - scale_vector_q16x8(exclude_padding, res_upper, id, 1, 2, - pool_size, upper_bound_w, upper_bound_h, - pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - upper_res = wrapper::vmovn(res_upper); - } - } - else - { - const q8x16_t max_data = wrapper::vmax(top_data, bottom_data); - lower_res = wrapper::vpmax(wrapper::vgetlow(max_data), wrapper::vgethigh(max_data)); - if(pool_stride_x == 1) - { - const q8x16_t max_data_shifted = wrapper::vext_1(max_data, max_data); - upper_res = wrapper::vpmax(wrapper::vgetlow(max_data_shifted), wrapper::vgethigh(max_data_shifted)); - } - } - - if(have_different_qinfo) - { - const auto requantized_output = vrequantize_pooling(lower_res, upper_res, requant_qinfo); - lower_res = wrapper::vgetlow(requantized_output); - upper_res = wrapper::vgethigh(requantized_output); - } - - // Store result - if(pool_stride_x == 1) - { - const q8x8x2_t res = { { lower_res, upper_res } }; - wrapper::vstore(reinterpret_cast(output.ptr()), res); - } - else - { - wrapper::vstore(reinterpret_cast(output.ptr()), lower_res); - } - }, - input, output); -} - -void NEPoolingLayerKernel::pooling3_f16_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - ARM_COMPUTE_UNUSED(pooling_type); - ARM_COMPUTE_UNUSED(exclude_padding); -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - Iterator input(_input, window_input); - Iterator output(_output, window); - - constexpr const int pool_size = 3; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top))); - const unsigned char *const input_middle_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1)); - const unsigned char *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 2)); - - execute_window_loop(window, [&](const Coordinates & id) - { - float16x4_t top_data = vld1_f16(reinterpret_cast(input_top_ptr + input.offset())); - float16x4_t middle_data = vld1_f16(reinterpret_cast(input_middle_ptr + input.offset())); - float16x4_t bottom_data = vld1_f16(reinterpret_cast(input_bottom_ptr + input.offset())); - float16x4_t res = {}; - - // Get power of 2 in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - top_data = vmul_f16(top_data, top_data); - middle_data = vmul_f16(middle_data, middle_data); - bottom_data = vmul_f16(bottom_data, bottom_data); - } - - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - const float16x4_t scale_v = vdup_n_f16(scale); - // Perform pooling - const float16x4_t sum_data = vadd_f16(vadd_f16(top_data, bottom_data), middle_data); - res = vpadd_f16(vset_lane_f16(0.f, sum_data, 3), sum_data); - res = vmul_f16(vpadd_f16(res, res), scale_v); - } - else - { - const float16x4_t max_data = vmax_f16(vmax_f16(top_data, bottom_data), middle_data); - res = vpmax_f16(vset_lane_f16(-std::numeric_limits::max(), max_data, 3), max_data); - res = vpmax_f16(res, res); - } - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - res = vinv_f16(vinvsqrt_f16(res)); - } - - *(reinterpret_cast(output.ptr())) = vget_lane_f16(res, 0); - }, - input, output); -#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - ARM_COMPUTE_UNUSED(window_input); - ARM_COMPUTE_UNUSED(window); - ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ -} - -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -template -inline typename std::enable_if::value, float32x2_t>::type -f16_to_f32(float16x4_t input) -{ - float32x2_t output = { static_cast(vget_lane_f16(input, 0)), static_cast(vget_lane_f16(input, 1)) }; - return output; -} -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - -template -inline typename std::enable_if::value, float32x2_t>::type -f16_to_f32(float32x2_t input) -{ - return input; -} - -template -void NEPoolingLayerKernel::pooling2_nchw_maxpool_indices(const Window &window_input, const Window &window) -{ - Iterator input(_input, window_input); - Iterator output(_output, window); - Iterator indices(_indices, window); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top))); - const uint8_t *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1)); - const int pad_left = _input->info()->padding().left; - const int pad_right = _input->info()->padding().right; - const int in_stride_y = static_cast(_input->info()->strides_in_bytes().y()); - - execute_window_loop(window, [&](const Coordinates & id) - { - auto top_data = wrapper::vload(reinterpret_cast(input_top_ptr + input.offset())); - auto bottom_data = wrapper::vload(reinterpret_cast(input_bottom_ptr + input.offset())); - float32x2_t top_data_f32 = f16_to_f32(top_data); - float32x2_t bottom_data_f32 = f16_to_f32(bottom_data); - - // Calculate max data, compare top first, then bottom, to make sue the first max is recorded. - const float32x2_t max_data_top = vpmax_f32(top_data_f32, top_data_f32); - const float32x2_t max_data_bottom = vpmax_f32(bottom_data_f32, bottom_data_f32); - const float32x2_t max_data = vmax_f32(max_data_top, max_data_bottom); - *(reinterpret_cast(output.ptr())) = static_cast(vget_lane_f32(max_data, 0)); - - // Calculate max data indice, which will be used in max unpool. - const uint32_t offset_base = offset_no_padding(input.offset(), id, *_input->info(), pool_stride_x, pool_stride_y); - const uint32_t offset_top = (uint32_t)(offset_base / sizeof(T)); - const uint32_t offset_bottom = offset_top + in_stride_y / sizeof(T) - pad_right - pad_left; - const uint32x2_t voffset_top = { offset_top, offset_top + 1u }; - const uint32x2_t voffset_bottom = { offset_bottom, offset_bottom + 1u }; - const uint32x2_t tmp_indices_top = vbsl_u32(vcge_f32(top_data_f32, vrev64_f32(top_data_f32)), voffset_top, vrev64_u32(voffset_top)); - const uint32x2_t tmp_indices_bottom = vbsl_u32(vcge_f32(bottom_data_f32, vrev64_f32(bottom_data_f32)), voffset_bottom, vrev64_u32(voffset_bottom)); - *(reinterpret_cast(indices.ptr())) = vget_lane_u32(vbsl_u32(vcge_f32(max_data_top, max_data_bottom), tmp_indices_top, tmp_indices_bottom), 0); - }, - input, output, indices); -} - -void NEPoolingLayerKernel::pooling2_f16_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - ARM_COMPUTE_UNUSED(pooling_type); - ARM_COMPUTE_UNUSED(exclude_padding); -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - if(pooling_type == PoolingType::MAX && _indices) - { - pooling2_nchw_maxpool_indices(window_input, window); - } - else - { - Iterator input(_input, window_input); - Iterator output(_output, window); - constexpr int pool_size = 2; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x, pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top))); - const unsigned char *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1)); - - execute_window_loop(window, [&](const Coordinates & id) - { - float16x4_t top_data = vld1_f16(reinterpret_cast(input_top_ptr + input.offset())); - float16x4_t bottom_data = vld1_f16(reinterpret_cast(input_bottom_ptr + input.offset())); - float16x4_t res = {}; - - // Get power of 2 in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - top_data = vmul_f16(top_data, top_data); - bottom_data = vmul_f16(bottom_data, bottom_data); - } - - if(pooling_type != PoolingType::MAX) - { - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - const float16x4_t scale_v = vdup_n_f16(scale); - - const float16x4_t sum_data = vadd_f16(top_data, bottom_data); - res = vmul_f16(vpadd_f16(sum_data, sum_data), scale_v); - } - else - { - const float16x4_t max_data = vmax_f16(top_data, bottom_data); - res = vpmax_f16(max_data, max_data); - } - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - res = vinv_f16(vinvsqrt_f16(res)); - } - - // Store result - *(reinterpret_cast(output.ptr())) = vget_lane_f16(res, 0); - }, - input, output); - } -#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - ARM_COMPUTE_UNUSED(window_input); - ARM_COMPUTE_UNUSED(window); - ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ -} - -template -void NEPoolingLayerKernel::pooling3_q8_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - Iterator input(_input, window_input); - Iterator output(_output, window); - - /** NEON vector types */ - using q8x8_t = typename wrapper::traits::neon_vector::type; - using q8x16_t = typename wrapper::traits::neon_vector::type; - using q8x8x2_t = typename std::conditional::value, uint8x8x2_t, int8x8x2_t>::type; - using q16_t = typename wrapper::traits::promote_t; - using q16x8_t = typename wrapper::traits::neon_vector::type; - using q16x8x2_t = typename wrapper::traits::neon_vector::type; - - constexpr int pool_size = 3; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - const UniformQuantizationInfo &input_qinfo = _input->info()->quantization_info().uniform(); - const UniformQuantizationInfo &output_qinfo = _output->info()->quantization_info().uniform(); - - const float requant_scale = output_qinfo.scale / input_qinfo.scale; - const int32_t requant_offset = output_qinfo.offset - static_cast(static_cast(input_qinfo.offset) / requant_scale); - const UniformQuantizationInfo requant_qinfo = UniformQuantizationInfo(requant_scale, requant_offset); - - const T *const input_top_ptr = reinterpret_cast(_input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top)))); - const T *const input_middle_ptr = reinterpret_cast(_input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1))); - const T *const input_bottom_ptr = reinterpret_cast(_input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 2))); - - execute_window_loop(window, [&](const Coordinates & id) - { - const auto top_data = wrapper::vloadq(input_top_ptr + input.offset()); - const auto middle_data = wrapper::vloadq(input_middle_ptr + input.offset()); - const auto bottom_data = wrapper::vloadq(input_bottom_ptr + input.offset()); - q8x8_t fres = {}; - q8x16_t fqres = {}; - - if(pooling_type == PoolingType::AVG) - { - // Convert data to u16 - const q16x8x2_t top_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(top_data)), wrapper::vmovl(wrapper::vgethigh(top_data)) } }; - const q16x8x2_t middle_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(middle_data)), wrapper::vmovl(wrapper::vgethigh(middle_data)) } }; - const q16x8x2_t bottom_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(bottom_data)), wrapper::vmovl(wrapper::vgethigh(bottom_data)) } }; - - // Calculate row sums - const q16x8x2_t vrsum = - { - { - wrapper::vadd(wrapper::vadd(top_data_q16.val[0], bottom_data_q16.val[0]), middle_data_q16.val[0]), - wrapper::vadd(wrapper::vadd(top_data_q16.val[1], bottom_data_q16.val[1]), middle_data_q16.val[1]), - } - }; - const q16x8x2_t vrsum_shifted_1 = - { - { - wrapper::vext_1(vrsum.val[0], vrsum.val[1]), - wrapper::vext_1(vrsum.val[1], vrsum.val[1]) - } - }; - const q16x8x2_t vrsum_shifted_2 = - { - { - wrapper::vext_2(vrsum.val[0], vrsum.val[1]), - wrapper::vext_2(vrsum.val[1], vrsum.val[1]) - } - }; - // Calculate final sum - q16x8x2_t final_sum = - { - { - wrapper::vadd(wrapper::vadd(vrsum.val[0], vrsum_shifted_1.val[0]), vrsum_shifted_2.val[0]), - wrapper::vadd(wrapper::vadd(vrsum.val[1], vrsum_shifted_1.val[1]), vrsum_shifted_2.val[1]), - } - }; - if(pool_stride_x == 2) - { - q16x8_t res = - { - wrapper::vgetlane(final_sum.val[0], 0), - wrapper::vgetlane(final_sum.val[0], 2), - wrapper::vgetlane(final_sum.val[0], 4), - wrapper::vgetlane(final_sum.val[0], 6), - wrapper::vgetlane(final_sum.val[1], 0), - wrapper::vgetlane(final_sum.val[1], 2), - wrapper::vgetlane(final_sum.val[1], 4), - wrapper::vgetlane(final_sum.val[1], 6), - }; - - scale_vector_q16x8(exclude_padding, res, id, 0, 1, - pool_size, upper_bound_w, upper_bound_h, - pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - fres = wrapper::vmovn(res); - } - else - { - // Scale lower result - scale_vector_q16x8(exclude_padding, final_sum.val[0], id, 0, 1, - pool_size, upper_bound_w, upper_bound_h, - pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - // Scale lower result - scale_vector_q16x8(exclude_padding, final_sum.val[1], id, 8, 1, - pool_size, upper_bound_w, upper_bound_h, - pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - fqres = wrapper::vcombine(wrapper::vmovn(final_sum.val[0]), wrapper::vmovn(final_sum.val[1])); - } - } - else - { - const q8x16_t max_data = wrapper::vmax(wrapper::vmax(top_data, bottom_data), middle_data); - const q8x16_t max_data_shift1 = wrapper::vext_1(max_data, max_data); - const q8x16_t max_data_shift2 = wrapper::vext_2(max_data, max_data); - const q8x16_t final_max = wrapper::vmax(wrapper::vmax(max_data, max_data_shift1), max_data_shift2); - - if(pool_stride_x == 2) - { - const q8x8x2_t table = { { wrapper::vgetlow(final_max), wrapper::vgethigh(final_max) } }; - static const q8x8_t lookup_val = { 0, 2, 4, 6, 8, 10, 12, 14 }; - fres = wrapper::vtbl(table, lookup_val); - } - else - { - fqres = final_max; - } - } - - // Store result - if(pool_stride_x == 1) - { - if(input_qinfo != output_qinfo) - { - fqres = vrequantize_pooling(wrapper::vgetlow(fqres), wrapper::vgethigh(fqres), requant_qinfo); - } - wrapper::vstore(reinterpret_cast(output.ptr()), fqres); - } - else - { - if(input_qinfo != output_qinfo) - { - fres = vrequantize_pooling(fres, requant_qinfo); - } - wrapper::vstore(reinterpret_cast(output.ptr()), fres); - } - }, - input, output); -} - -void NEPoolingLayerKernel::poolingMxN_f16_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - ARM_COMPUTE_UNUSED(pooling_type); - ARM_COMPUTE_UNUSED(exclude_padding); -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - Iterator input(_input, window_input); - Iterator output(_output, window); - - const int pool_size_x = _pool_info.is_global_pooling ? _input->info()->tensor_shape().x() : _pool_info.pool_size.width; - const int pool_size_y = _pool_info.is_global_pooling ? _input->info()->tensor_shape().y() : _pool_info.pool_size.height; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - execute_window_loop(window, [&](const Coordinates & id) - { - float16_t res = 0.0f; - float16x8_t vres = vdupq_n_f16(0.0f); - - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - - // Perform pooling - - for(int y = 0; y < pool_size_y; ++y) - { - int x = 0; - for(; x <= (pool_size_x - 8); x += 8) - { - const float16x8_t data = vld1q_f16(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + - (y - pool_pad_top) * static_cast(_input->info()->strides_in_bytes().y()))); - - // Get power of 2 in case of l2 pooling and accumulate - if(pooling_type == PoolingType::L2) - { - vres = vaddq_f16(vres, vmulq_f16(data, data)); - } - else - { - vres = vaddq_f16(vres, data); - } - } - - // Leftover for loop - for(; x < pool_size_x; ++x) - { - float16_t data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) - + (y - pool_pad_top) * static_cast(_input->info()->strides_in_bytes().y()))); - - // Get power of 2 in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - data *= data; - } - - res += data; - } - } - - // Reduction - float16x4_t tmp = vpadd_f16(vget_high_f16(vres), vget_low_f16(vres)); - res += vget_lane_f16(tmp, 0); - res += vget_lane_f16(tmp, 1); - res += vget_lane_f16(tmp, 2); - res += vget_lane_f16(tmp, 3); - - // Divide by scale - res *= scale; - } - else - { - float16x8_t vres = vdupq_n_f16(std::numeric_limits::lowest()); - res = std::numeric_limits::lowest(); - - for(int y = 0; y < pool_size_y; ++y) - { - int x = 0; - for(; x <= (pool_size_x - 8); x += 8) - { - const float16x8_t data = vld1q_f16(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + - (y - pool_pad_top) * static_cast(_input->info()->strides_in_bytes().y()))); - vres = vmaxq_f16(vres, data); - } - - // Leftover for loop - for(; x < pool_size_x; ++x) - { - const float16_t data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) - + (y - pool_pad_top) * static_cast(_input->info()->strides_in_bytes().y()))); - res = std::max(res, data); - } - } - - float16x4_t tmp = vpmax_f16(vget_high_f16(vres), vget_low_f16(vres)); - res = std::max(res, vget_lane_f16(tmp, 0)); - res = std::max(res, vget_lane_f16(tmp, 1)); - res = std::max(res, vget_lane_f16(tmp, 2)); - res = std::max(res, vget_lane_f16(tmp, 3)); - } - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - res = std::sqrt(res); - } - - // Store result - *(reinterpret_cast(output.ptr())) = res; - }, - input, output); - -#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - ARM_COMPUTE_UNUSED(window_input); - ARM_COMPUTE_UNUSED(window); - ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ -} - -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -void NEPoolingLayerKernel::pooling2_f16_nhwc_maxpool_indices(const Window &window_input, const Window &window) -{ - const int window_start_x = window.x().start(); - const int window_end_x = window.x().end(); - const int window_step_x = 8; - - Window window_out = window; - window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input(_input, window_input); - Iterator output(_output, window_out); - Iterator indices(_indices, window_out); - - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - - const int pad_right = _input->info()->padding().right; - const int in_stride_y = static_cast(_input->info()->strides_in_bytes().y()); - const int in_stride_z = static_cast(_input->info()->strides_in_bytes().z()); - - execute_window_loop(window_out, [&](const Coordinates & id) - { - const int idx_width = id.y() * pool_stride_x; - const int idx_height = id.z() * pool_stride_y; - const int pool_limit_y = pool_pad_top - idx_height; - const int pool_limit_x = pool_pad_left - idx_width; - - const int pool_start_y = std::max(0, window_input.z().start() + pool_limit_y); - const int pool_start_x = std::max(0, window_input.y().start() + pool_limit_x); - const int in_x0_offset = (pool_start_x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (pool_start_y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z()); - const int in_x1_offset = (pool_start_x + 1 - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (pool_start_y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z()); - const int in_x2_offset = (pool_start_x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (pool_start_y + 1 - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z()); - const int in_x3_offset = (pool_start_x + 1 - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (pool_start_y + 1 - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z()); - - int x_off = window_start_x; - for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) - { - const auto in_x0_ptr = reinterpret_cast(input.ptr() + in_x0_offset) + x_off; - const auto in_x1_ptr = reinterpret_cast(input.ptr() + in_x1_offset) + x_off; - const auto in_x2_ptr = reinterpret_cast(input.ptr() + in_x2_offset) + x_off; - const auto in_x3_ptr = reinterpret_cast(input.ptr() + in_x3_offset) + x_off; - const auto v_x0 = vld1q_f16(in_x0_ptr); - const auto v_x1 = vld1q_f16(in_x1_ptr); - const auto v_x2 = vld1q_f16(in_x2_ptr); - const auto v_x3 = vld1q_f16(in_x3_ptr); - float16x8_t vres = vmaxq_f16(vmaxq_f16(v_x2, v_x3), vmaxq_f16(v_x0, v_x1)); - // Store result - vst1q_f16(reinterpret_cast(output.ptr()) + x_off, vres); - - const uint32_t offset_base = offset_no_padding(input.offset(), id, *_input->info(), pool_stride_x, pool_stride_y); - const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float16_t) + x_off; - const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float16_t) - pad_right; - const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float16_t) - pad_right * _input->info()->tensor_shape()[1]; - const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float16_t) - pad_right; - const uint32x4_t voffset_x0_0 = { offset_x0, offset_x0 + 1, offset_x0 + 2, offset_x0 + 3 }; - const uint32x4_t voffset_x0_1 = { offset_x0 + 4, offset_x0 + 5, offset_x0 + 6, offset_x0 + 7 }; - const uint16x8_t voffset_x0 = vcombine_u16(vmovn_u32(voffset_x0_0), vmovn_u32(voffset_x0_1)); - const uint32x4_t voffset_x1_0 = { offset_x1, offset_x1 + 1, offset_x1 + 2, offset_x1 + 3 }; - const uint32x4_t voffset_x1_1 = { offset_x1 + 4, offset_x1 + 5, offset_x1 + 6, offset_x1 + 7 }; - const uint16x8_t voffset_x1 = vcombine_u16(vmovn_u32(voffset_x1_0), vmovn_u32(voffset_x1_1)); - const uint32x4_t voffset_x2_0 = { offset_x2, offset_x2 + 1, offset_x2 + 2, offset_x2 + 3 }; - const uint32x4_t voffset_x2_1 = { offset_x2 + 4, offset_x2 + 5, offset_x2 + 6, offset_x2 + 7 }; - const uint16x8_t voffset_x2 = vcombine_u16(vmovn_u32(voffset_x2_0), vmovn_u32(voffset_x2_1)); - const uint32x4_t voffset_x3_0 = { offset_x3, offset_x3 + 1, offset_x3 + 2, offset_x3 + 3 }; - const uint32x4_t voffset_x3_1 = { offset_x3 + 4, offset_x3 + 5, offset_x3 + 6, offset_x3 + 7 }; - const uint16x8_t voffset_x3 = vcombine_u16(vmovn_u32(voffset_x3_0), vmovn_u32(voffset_x3_1)); - const uint16x8_t tmp_indices0 = vbslq_u16(vcgeq_f16(v_x0, v_x1), voffset_x0, voffset_x1); - const uint16x8_t tmp_indices1 = vbslq_u16(vcgeq_f16(v_x2, v_x3), voffset_x2, voffset_x3); - const uint16x8_t tmp_indices2 = vbslq_u16(vcgeq_f16(vmaxq_f16(v_x0, v_x1), vmaxq_f16(v_x2, v_x3)), tmp_indices0, tmp_indices1); - const uint32x4_t tmp_indeces3_0 = vmovl_u16(vget_low_u16(tmp_indices2)); - const uint32x4_t tmp_indeces3_1 = vmovl_u16(vget_high_u16(tmp_indices2)); - // Store indicies - vst1q_u32(reinterpret_cast(indices.ptr()) + x_off, tmp_indeces3_0); - vst1q_u32(reinterpret_cast(indices.ptr() + 16) + x_off, tmp_indeces3_1); - } - - // Left-overs loop - for(; x_off < window_end_x; ++x_off) - { - const auto x0 = *(reinterpret_cast(input.ptr() + in_x0_offset) + x_off); - const auto x1 = *(reinterpret_cast(input.ptr() + in_x1_offset) + x_off); - const auto x2 = *(reinterpret_cast(input.ptr() + in_x2_offset) + x_off); - const auto x3 = *(reinterpret_cast(input.ptr() + in_x3_offset) + x_off); - float16_t res = std::max(std::max(x2, x3), std::max(x0, x1)); - - // Store result - *(reinterpret_cast(output.ptr()) + x_off) = res; - - const uint32_t offset_base = offset_no_padding(input.offset(), id, *_input->info(), pool_stride_x, pool_stride_y); - const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float16_t) + x_off; - const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float16_t) - pad_right; - const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float16_t) - pad_right * _input->info()->tensor_shape()[1]; - const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float16_t) - pad_right; - const uint32_t tmp_idx0 = (x0 >= x1) ? offset_x0 : offset_x1; - const uint32_t tmp_idx1 = (x2 >= x3) ? offset_x2 : offset_x3; - const uint32_t tmp_idx2 = (std::max(x0, x1) >= std::max(x2, x3)) ? tmp_idx0 : tmp_idx1; - - // Store indices - *(reinterpret_cast(indices.ptr()) + x_off) = tmp_idx2; - } - }, - input, output, indices); -} -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - -void NEPoolingLayerKernel::poolingMxN_f16_nhwc(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - ARM_COMPUTE_UNUSED(pooling_type); - ARM_COMPUTE_UNUSED(exclude_padding); -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - if(_pool_info.pool_size == Size2D(2, 2) && pooling_type == PoolingType::MAX && _indices) - { - pooling2_f16_nhwc_maxpool_indices(window_input, window); - } - const int window_start_x = window.x().start(); - const int window_end_x = window.x().end(); - const int window_step_x = 8; - - Window window_out = window; - window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input(_input, window_input); - Iterator output(_output, window_out); - - const int pool_size_x = _pool_info.is_global_pooling ? _input->info()->tensor_shape().y() : _pool_info.pool_size.width; - const int pool_size_y = _pool_info.is_global_pooling ? _input->info()->tensor_shape().z() : _pool_info.pool_size.height; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _input->info()->dimension(2) + (exclude_padding ? 0 : pool_pad_bottom); - - float16x8_t vres; - - execute_window_loop(window_out, [&](const Coordinates & id) - { - const int idx_width = id.y() * pool_stride_x; - const int idx_height = id.z() * pool_stride_y; - const int pool_limit_y = pool_pad_top - idx_height; - const int pool_limit_x = pool_pad_left - idx_width; - - const int pool_start_y = std::max(0, window_input.z().start() + pool_limit_y); - const int pool_end_y = std::min(pool_size_y, window_input.z().end() + pool_limit_y); - const int pool_start_x = std::max(0, window_input.y().start() + pool_limit_x); - const int pool_end_x = std::min(pool_size_x, window_input.y().end() + pool_limit_x); - - int x_off = window_start_x; - for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) - { - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, - pool_stride_y); - const float16x8_t scale_v = vdupq_n_f16(scale); - - // Perform pooling - vres = vdupq_n_f16(0.0f); - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const float16x8_t data = vld1q_f16(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + - (y - pool_pad_top) * static_cast(_input->info()->strides_in_bytes().z())) + x_off); - - // Get power of 2 in case of l2 pooling and accumulate - if(pooling_type == PoolingType::L2) - { - vres = vaddq_f16(vres, vmulq_f16(data, data)); - } - else - { - vres = vaddq_f16(vres, data); - } - } - } - // Divide by scale - vres = vmulq_f16(vres, scale_v); - } - else - { - vres = vdupq_n_f16(std::numeric_limits::lowest()); - - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const float16x8_t data = vld1q_f16(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + - (y - pool_pad_top) * static_cast(_input->info()->strides_in_bytes().z())) + x_off); - vres = vmaxq_f16(vres, data); - } - } - } - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - float16x8_t sqrt_reciprocal = vrsqrteq_f16(vres); - vres = vmulq_f16(vres, vmulq_f16(vrsqrtsq_f16(vmulq_f16(vres, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal)); - } - - // Store result - vst1q_f16(reinterpret_cast(output.ptr()) + x_off, vres); - } - - // Left-overs loop - for(; x_off < window_end_x; ++x_off) - { - float16_t res = 0.0f; - - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - const float16_t scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, - pool_stride_y); - - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const float data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z())) + x_off); - - // Get power of 2 in case of l2 pooling and accumulate - if(pooling_type == PoolingType::L2) - { - res += data * data; - } - else - { - res += data; - } - } - } - - // Divide by scale - res *= scale; - } - else - { - res = std::numeric_limits::lowest(); - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const float16_t data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z())) + x_off); - res = std::max(res, data); - } - } - } - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - res = std::sqrt(res); - } - - // Store result - *(reinterpret_cast(output.ptr()) + x_off) = res; - } - }, - input, output); - -#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - ARM_COMPUTE_UNUSED(window_input); - ARM_COMPUTE_UNUSED(window); - ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ -} - -void NEPoolingLayerKernel::poolingMxN_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - Iterator input(_input, window_input); - Iterator output(_output, window); - - const int pool_size_x = _pool_info.is_global_pooling ? _input->info()->tensor_shape().x() : _pool_info.pool_size.width; - const int pool_size_y = _pool_info.is_global_pooling ? _input->info()->tensor_shape().y() : _pool_info.pool_size.height; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - execute_window_loop(window, [&](const Coordinates & id) - { - float res = 0.0f; - - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - - // Perform pooling - float32x4_t vres = vdupq_n_f32(0.0f); - - for(int y = 0; y < pool_size_y; ++y) - { - int x = 0; - for(; x <= (pool_size_x - 4); x += 4) - { - const float32x4_t data = vld1q_f32(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().y()))); - - // Get power of 2 in case of l2 pooling and accumulate - if(pooling_type == PoolingType::L2) - { - vres = vmlaq_f32(vres, data, data); - } - else - { - vres = vaddq_f32(vres, data); - } - } - - // Leftover for loop - for(; x < pool_size_x; ++x) - { - float data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().y()))); - - // Get power of 2 in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - data *= data; - } - - res += data; - } - } - -#if defined(__aarch64__) - // Reduction operation available on 64 bit architectures only - res += vaddvq_f32(vres); -#else // __aarch64__ - // Reduction - float32x2_t tmp = vpadd_f32(vget_high_f32(vres), vget_low_f32(vres)); - tmp = vpadd_f32(tmp, tmp); - - res += vget_lane_f32(tmp, 0); -#endif // __aarch64__ - // Divide by scale - res *= scale; - } - else - { - float32x4_t vres = vdupq_n_f32(std::numeric_limits::lowest()); - res = std::numeric_limits::lowest(); - - for(int y = 0; y < pool_size_y; ++y) - { - int x = 0; - for(; x <= (pool_size_x - 4); x += 4) - { - const float32x4_t data = vld1q_f32(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().y()))); - vres = vmaxq_f32(vres, data); - } - - // Leftover for loop - for(; x < pool_size_x; ++x) - { - const float data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().y()))); - res = std::max(res, data); - } - } -#if defined(__aarch64__) - // Reduction operation available on 64 bit architectures only - res = std::max(vmaxvq_f32(vres), res); -#else // __aarch64__ - float32x2_t tmp = vpmax_f32(vget_high_f32(vres), vget_low_f32(vres)); - tmp = vpmax_f32(tmp, tmp); - - res = std::max(res, vget_lane_f32(tmp, 0)); -#endif // __aarch64__ - } - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - res = std::sqrt(res); - } - - // Store result - *(reinterpret_cast(output.ptr())) = res; - }, - input, output); -} - -void NEPoolingLayerKernel::pooling2_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, - bool exclude_padding) -{ - if(pooling_type == PoolingType::MAX && _indices) - { - pooling2_nchw_maxpool_indices(window_input, window); - } - else - { - Iterator input(_input, window_input); - Iterator output(_output, window); - constexpr int pool_size = 2; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top))); - const uint8_t *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1)); - - execute_window_loop(window, [&](const Coordinates & id) - { - const auto in_top_ptr = reinterpret_cast(input_top_ptr + input.offset()); - const auto in_bottom_ptr = reinterpret_cast(input_bottom_ptr + input.offset()); - float32x2_t top_data = vld1_f32(in_top_ptr); - float32x2_t bottom_data = vld1_f32(in_bottom_ptr); - float32x2_t res = {}; - float final_res = 0; - // Get power of 2 in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - top_data = vmul_f32(top_data, top_data); - bottom_data = vmul_f32(bottom_data, bottom_data); - } - - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - const float32x2_t scale_v = vdup_n_f32(scale); - - // Perform pooling - const float32x2_t sum_data = vadd_f32(top_data, bottom_data); - res = vmul_f32(vpadd_f32(sum_data, sum_data), scale_v); - } - else - { - const float32x2_t max_data = vmax_f32(top_data, bottom_data); - res = vpmax_f32(max_data, max_data); - } - final_res = vget_lane_f32(res, 0); - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - final_res = sqrt(final_res); - } - - // Store result - *(reinterpret_cast(output.ptr())) = final_res; - }, - input, output); - } -} - -void NEPoolingLayerKernel::pooling3_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - Iterator input(_input, window_input); - Iterator output(_output, window); - - constexpr const int pool_size = 3; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top))); - const uint8_t *const input_middle_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1)); - const uint8_t *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 2)); - - execute_window_loop(window, [&](const Coordinates & id) - { - float32x4_t top_data = vld1q_f32(reinterpret_cast(input_top_ptr + input.offset())); - float32x4_t middle_data = vld1q_f32(reinterpret_cast(input_middle_ptr + input.offset())); - float32x4_t bottom_data = vld1q_f32(reinterpret_cast(input_bottom_ptr + input.offset())); - float32x2_t res = {}; - float final_res = 0; - - // Get power of 2 in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - top_data = vmulq_f32(top_data, top_data); - middle_data = vmulq_f32(middle_data, middle_data); - bottom_data = vmulq_f32(bottom_data, bottom_data); - } - - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - const float32x2_t scale_v = vdup_n_f32(scale); - - // Perform pooling - const float32x4_t sum_data = vaddq_f32(vaddq_f32(top_data, bottom_data), middle_data); - res = vpadd_f32(vget_high_f32(vsetq_lane_f32(0.f, sum_data, 3)), vget_low_f32(sum_data)); - res = vmul_f32(vpadd_f32(res, res), scale_v); - } - else - { - const float32x4_t max_data = vmaxq_f32(vmaxq_f32(top_data, bottom_data), middle_data); - res = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits::max(), max_data, 3)), vget_low_f32(max_data)); - res = vpmax_f32(res, res); - } - final_res = vget_lane_f32(res, 0); - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - final_res = sqrt(final_res); - } - - // Store result - *(reinterpret_cast(output.ptr())) = final_res; - }, - input, output); -} - -void NEPoolingLayerKernel::pooling7_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - Iterator input(_input, window_input); - Iterator output(_output, window); - - constexpr const int pool_size = 7; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - std::array input_ptrs{ {} }; - for(int i = 0; i < pool_size; ++i) - { - input_ptrs[i] = _input->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + i)); - } - - execute_window_loop(window, [&](const Coordinates & id) - { - float32x2_t res = {}; - float final_res = 0.f; - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - const float32x2_t scale_v = vdup_n_f32(scale); - - // Perform pooling - float32x4x2_t data = vld2q_f32(reinterpret_cast(input_ptrs[0] + input.offset())); - // Get power of 2 in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - data.val[0] = vmulq_f32(data.val[0], data.val[0]); - data.val[1] = vmulq_f32(data.val[1], data.val[1]); - } - float32x4_t sum_data = vaddq_f32(data.val[0], vsetq_lane_f32(0.f, data.val[1], 3)); - for(int i = 1; i < pool_size; ++i) - { - data = vld2q_f32(reinterpret_cast(input_ptrs[i] + input.offset())); - // Get power of 2 in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - data.val[0] = vmulq_f32(data.val[0], data.val[0]); - data.val[1] = vmulq_f32(data.val[1], data.val[1]); - } - sum_data = vaddq_f32(sum_data, data.val[0]); - sum_data = vaddq_f32(sum_data, vsetq_lane_f32(0.f, data.val[1], 3)); - } - res = vpadd_f32(vget_high_f32(sum_data), vget_low_f32(sum_data)); - res = vmul_f32(vpadd_f32(res, res), scale_v); - } - else - { - float32x4x2_t max_data = vld2q_f32(reinterpret_cast(input_ptrs[0] + input.offset())); - for(int i = 1; i < pool_size; ++i) - { - const float32x4x2_t data = vld2q_f32(reinterpret_cast(input_ptrs[i] + input.offset())); - max_data = vmax2q_f32(max_data, data); - } - res = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits::max(), max_data.val[1], 3)), vget_low_f32(max_data.val[1])); - res = vpmax_f32(res, vpmax_f32(vget_high_f32(max_data.val[0]), vget_low_f32(max_data.val[0]))); - res = vpmax_f32(res, res); - } - final_res = vget_lane_f32(res, 0); - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - final_res = sqrt(final_res); - } - - // Store result - *(reinterpret_cast(output.ptr())) = final_res; - }, - input, output); -} - -void NEPoolingLayerKernel::poolingMxN_f32_nhwc(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - if(_pool_info.pool_size == Size2D(2, 2) && pooling_type == PoolingType::MAX && _indices) - { - pooling2_f32_nhwc_maxpool_indices(window_input, window); - } - else - { - const int window_start_x = window.x().start(); - const int window_end_x = window.x().end(); - const int window_step_x = 4; - - Window window_out = window; - window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input(_input, window_input); - Iterator output(_output, window_out); - - const int pool_size_x = _pool_info.is_global_pooling ? _input->info()->tensor_shape().y() : _pool_info.pool_size.width; - const int pool_size_y = _pool_info.is_global_pooling ? _input->info()->tensor_shape().z() : _pool_info.pool_size.height; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _input->info()->dimension(2) + (exclude_padding ? 0 : pool_pad_bottom); - - float32x4_t vres; - - execute_window_loop(window_out, [&](const Coordinates & id) - { - const int idx_width = id.y() * pool_stride_x; - const int idx_height = id.z() * pool_stride_y; - const int pool_limit_y = pool_pad_top - idx_height; - const int pool_limit_x = pool_pad_left - idx_width; - - const int pool_start_y = std::max(0, window_input.z().start() + pool_limit_y); - const int pool_end_y = std::min(pool_size_y, window_input.z().end() + pool_limit_y); - const int pool_start_x = std::max(0, window_input.y().start() + pool_limit_x); - const int pool_end_x = std::min(pool_size_x, window_input.y().end() + pool_limit_x); - - int x_off = window_start_x; - for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) - { - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, - pool_stride_y); - const float32x4_t scale_v = vdupq_n_f32(scale); - - // Perform pooling - vres = vdupq_n_f32(0.0f); - - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const float32x4_t data = vld1q_f32(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z())) + x_off); - - // Get power of 2 in case of l2 pooling and accumulate - if(pooling_type == PoolingType::L2) - { - vres = vmlaq_f32(vres, data, data); - } - else - { - vres = vaddq_f32(vres, data); - } - } - } - // Divide by scale - vres = vmulq_f32(vres, scale_v); - } - else - { - vres = vdupq_n_f32(std::numeric_limits::lowest()); - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const float32x4_t data = vld1q_f32(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z())) + x_off); - vres = vmaxq_f32(vres, data); - } - } - } - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - float32x4_t l2_res = { static_cast(sqrt(vgetq_lane_f32(vres, 0))), - static_cast(sqrt(vgetq_lane_f32(vres, 1))), - static_cast(sqrt(vgetq_lane_f32(vres, 2))), - static_cast(sqrt(vgetq_lane_f32(vres, 3))) - }; - vres = l2_res; - } - - // Store result - vst1q_f32(reinterpret_cast(output.ptr()) + x_off, vres); - } - - // Left-overs loop - for(; x_off < window_end_x; ++x_off) - { - float res = 0.0f; - - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, - pool_stride_y); - - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const float data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z())) + x_off); - - // Get power of 2 in case of l2 pooling and accumulate - if(pooling_type == PoolingType::L2) - { - res += data * data; - } - else - { - res += data; - } - } - } - - // Divide by scale - res *= scale; - } - else - { - res = std::numeric_limits::lowest(); - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const float data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z())) + x_off); - res = std::max(res, data); - } - } - } - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - res = std::sqrt(res); - } - - // Store result - *(reinterpret_cast(output.ptr()) + x_off) = res; - } - }, - input, output); - } -} - -void NEPoolingLayerKernel::pooling2_f32_nhwc_maxpool_indices(const Window &window_input, const Window &window) -{ - const int window_start_x = window.x().start(); - const int window_end_x = window.x().end(); - const int window_step_x = 4; - - Window window_out = window; - window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input(_input, window_input); - Iterator output(_output, window_out); - Iterator indices(_indices, window_out); - - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - - float32x4_t vres; - float res; - - const int pad_right = _input->info()->padding().right; - const int in_stride_y = static_cast(_input->info()->strides_in_bytes().y()); - const int in_stride_z = static_cast(_input->info()->strides_in_bytes().z()); - - execute_window_loop(window_out, [&](const Coordinates & id) - { - const int idx_width = id.y() * pool_stride_x; - const int idx_height = id.z() * pool_stride_y; - const int pool_limit_y = pool_pad_top - idx_height; - const int pool_limit_x = pool_pad_left - idx_width; - - const int pool_start_y = std::max(0, window_input.z().start() + pool_limit_y); - const int pool_start_x = std::max(0, window_input.y().start() + pool_limit_x); - - const int in_x0_offset = (pool_start_x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (pool_start_y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z()); - const int in_x1_offset = (pool_start_x + 1 - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (pool_start_y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z()); - const int in_x2_offset = (pool_start_x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (pool_start_y + 1 - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z()); - const int in_x3_offset = (pool_start_x + 1 - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (pool_start_y + 1 - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z()); - - int x_off = window_start_x; - for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) - { - const auto in_x0_ptr = reinterpret_cast(input.ptr() + in_x0_offset); - const auto in_x1_ptr = reinterpret_cast(input.ptr() + in_x1_offset); - const auto in_x2_ptr = reinterpret_cast(input.ptr() + in_x2_offset); - const auto in_x3_ptr = reinterpret_cast(input.ptr() + in_x3_offset); - const auto v_x0 = vld1q_f32(in_x0_ptr + x_off); - const auto v_x1 = vld1q_f32(in_x1_ptr + x_off); - const auto v_x2 = vld1q_f32(in_x2_ptr + x_off); - const auto v_x3 = vld1q_f32(in_x3_ptr + x_off); - vres = vmaxq_f32(vmaxq_f32(v_x2, v_x3), vmaxq_f32(v_x0, v_x1)); - // Store result - vst1q_f32(reinterpret_cast(output.ptr()) + x_off, vres); - - const uint32_t offset_base = offset_no_padding(input.offset(), id, *_input->info(), pool_stride_x, pool_stride_y); - const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float) + x_off; - const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float) - pad_right; - const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float) - pad_right * _input->info()->tensor_shape()[1]; - const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float) - pad_right; - const uint32x4_t voffset_x0 = { offset_x0, offset_x0 + 1, offset_x0 + 2, offset_x0 + 3 }; - const uint32x4_t voffset_x1 = { offset_x1, offset_x1 + 1, offset_x1 + 2, offset_x1 + 3 }; - const uint32x4_t voffset_x2 = { offset_x2, offset_x2 + 1, offset_x2 + 2, offset_x2 + 3 }; - const uint32x4_t voffset_x3 = { offset_x3, offset_x3 + 1, offset_x3 + 2, offset_x3 + 3 }; - const uint32x4_t tmp_indices0 = vbslq_u32(vcgeq_f32(v_x0, v_x1), voffset_x0, voffset_x1); - const uint32x4_t tmp_indices1 = vbslq_u32(vcgeq_f32(v_x2, v_x3), voffset_x2, voffset_x3); - const uint32x4_t tmp_indices2 = vbslq_u32(vcgeq_f32(vmaxq_f32(v_x0, v_x1), vmaxq_f32(v_x2, v_x3)), tmp_indices0, tmp_indices1); - - // Store indices - vst1q_u32(reinterpret_cast(indices.ptr()) + x_off, tmp_indices2); - } - - // Left-overs loop - for(; x_off < window_end_x; ++x_off) - { - const auto x0 = *(reinterpret_cast(input.ptr() + in_x0_offset) + x_off); - const auto x1 = *(reinterpret_cast(input.ptr() + in_x1_offset) + x_off); - const auto x2 = *(reinterpret_cast(input.ptr() + in_x2_offset) + x_off); - const auto x3 = *(reinterpret_cast(input.ptr() + in_x3_offset) + x_off); - res = std::max(std::max(x2, x3), std::max(x0, x1)); - - // Store result - *(reinterpret_cast(output.ptr()) + x_off) = res; - - const uint32_t offset_base = offset_no_padding(input.offset(), id, *_input->info(), pool_stride_x, pool_stride_y); - const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float) + x_off; - const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float) - pad_right; - const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float) - pad_right * _input->info()->tensor_shape()[1]; - const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float) - pad_right; - const uint32_t tmp_idx0 = (x0 >= x1) ? offset_x0 : offset_x1; - const uint32_t tmp_idx1 = (x2 >= x3) ? offset_x2 : offset_x3; - const uint32_t tmp_idx2 = (std::max(x0, x1) >= std::max(x2, x3)) ? tmp_idx0 : tmp_idx1; - - // Store indices - *(reinterpret_cast(indices.ptr()) + x_off) = tmp_idx2; - } - }, - input, output, indices); -} - -template -void NEPoolingLayerKernel::poolingMxN_q8_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - Iterator input(_input, window_input); - Iterator output(_output, window); - - /** NEON vector types */ - using q8x8_t = typename wrapper::traits::neon_vector::type; - using q16_t = typename wrapper::traits::promote_t; - using q16x8_t = typename wrapper::traits::neon_vector::type; - using q32_t = typename wrapper::traits::promote_t; - using q32x4_t = typename wrapper::traits::neon_vector::type; - - const int pool_size_x = _pool_info.is_global_pooling ? _input->info()->tensor_shape().x() : _pool_info.pool_size.width; - const int pool_size_y = _pool_info.is_global_pooling ? _input->info()->tensor_shape().y() : _pool_info.pool_size.height; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - const UniformQuantizationInfo &input_qinfo = _input->info()->quantization_info().uniform(); - const UniformQuantizationInfo &output_qinfo = _output->info()->quantization_info().uniform(); - - execute_window_loop(window, [&](const Coordinates & id) - { - T res = std::numeric_limits::min(); - - if(pooling_type != PoolingType::MAX) - { - q32x4_t vres = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); - q32_t sres = 0; - - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - - // Perform pooling - for(int y = 0; y < pool_size_y; ++y) - { - int x = 0; - for(; x <= (pool_size_x - 8); x += 8) - { - const q8x8_t data = wrapper::vload(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().y()))); - - const q16x8_t data_q16 = wrapper::vmovl(data); - vres = wrapper::vadd(vres, wrapper::vaddl(wrapper::vgethigh(data_q16), wrapper::vgetlow(data_q16))); - } - - // Leftover for loop - for(; x < pool_size_x; ++x) - { - T data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().y()))); - sres += data; - } - } - - // Reduction - const auto tmp = wrapper::vpadd(wrapper::vgethigh(vres), wrapper::vgetlow(vres)); - sres += wrapper::vgetlane(tmp, 0) + wrapper::vgetlane(tmp, 1); - - // Divide by scale - res = static_cast(support::cpp11::round(sres * scale)); - } - else - { - q8x8_t vres = wrapper::vdup_n(std::numeric_limits::min(), wrapper::traits::vector_64_tag{}); - - for(int y = 0; y < pool_size_y; ++y) - { - int x = 0; - for(; x <= (pool_size_x - 8); x += 8) - { - const q8x8_t data = wrapper::vload(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().y()))); - vres = wrapper::vmax(vres, data); - } - // Leftover for loop - for(; x < pool_size_x; ++x) - { - const T data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().y()))); - res = std::max(res, data); - } - } - - // Reduce max - vres = wrapper::vpmax(vres, vres); - vres = wrapper::vpmax(vres, vres); - vres = wrapper::vpmax(vres, vres); - - // Get max value - res = std::max(res, wrapper::vgetlane(vres, 0)); - } - // Store result - res = (input_qinfo != output_qinfo) ? Qasymm8QuantizationHelper::quantize(Qasymm8QuantizationHelper::dequantize(res, input_qinfo), output_qinfo) : res; - *(reinterpret_cast(output.ptr())) = res; - }, - input, output); -} - -template -void NEPoolingLayerKernel::poolingMxN_q8_nhwc(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - const int window_start_x = window.x().start(); - const int window_end_x = window.x().end(); - const int window_step_x = 16; - const int window_half_step_x = window_step_x / 2; - - Window window_out = window; - window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator input(_input, window_input); - Iterator output(_output, window_out); - - using q8x8_t = typename wrapper::traits::neon_vector::type; - using q8x16_t = typename wrapper::traits::neon_vector::type; - using q16_t = typename wrapper::traits::promote_t; - using q16x8_t = typename wrapper::traits::neon_vector::type; - using q32_t = typename wrapper::traits::promote_t; - using q32x4_t = typename wrapper::traits::neon_vector::type; - - const int pool_size_x = _pool_info.is_global_pooling ? _input->info()->tensor_shape().y() : _pool_info.pool_size.width; - const int pool_size_y = _pool_info.is_global_pooling ? _input->info()->tensor_shape().z() : _pool_info.pool_size.height; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _input->info()->dimension(2) + (exclude_padding ? 0 : pool_pad_bottom); - - const float32x4_t half_scale_v = vdupq_n_f32(0.5f); - const UniformQuantizationInfo input_qinfo = _input->info()->quantization_info().uniform(); - const UniformQuantizationInfo output_qinfo = _output->info()->quantization_info().uniform(); - - const float quant_rescale = output_qinfo.scale / input_qinfo.scale; - // "new_offset" doesn't have to consider the "half_scale_v" in its computation - // With a requantization performed in a single step there won't be uncertainties introduced - const int32_t new_offset = output_qinfo.offset - static_cast(static_cast(input_qinfo.offset) / quant_rescale); - - const float requant_scale = output_qinfo.scale / input_qinfo.scale; - const int32_t requant_offset = output_qinfo.offset - static_cast(static_cast(input_qinfo.offset) / requant_scale); - const UniformQuantizationInfo requant_qinfo = UniformQuantizationInfo(requant_scale, requant_offset); - - execute_window_loop(window_out, [&](const Coordinates & id) - { - const int idx_width = id.y() * pool_stride_x; - const int idx_height = id.z() * pool_stride_y; - const int pool_limit_y = pool_pad_top - idx_height; - const int pool_limit_x = pool_pad_left - idx_width; - - const int pool_start_y = std::max(0, window_input.z().start() + pool_limit_y); - const int pool_end_y = std::min(pool_size_y, window_input.z().end() + pool_limit_y); - const int pool_start_x = std::max(0, window_input.y().start() + pool_limit_x); - const int pool_end_x = std::min(pool_size_x, window_input.y().end() + pool_limit_x); - - int x_off = window_start_x; - for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) - { - if(pooling_type != PoolingType::MAX) - { - q32x4_t vres1 = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); - q32x4_t vres2 = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); - q32x4_t vres3 = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); - q32x4_t vres4 = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); - - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, - pool_stride_y); - - // Perform pooling - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const q8x16_t data = wrapper::vloadq(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z())) + x_off); - - const q16x8_t data_q16 = wrapper::vmovl(wrapper::vgetlow(data)); - const q16x8_t data2_q16 = wrapper::vmovl(wrapper::vgethigh(data)); - vres1 = wrapper::vadd(vres1, wrapper::vmovl(wrapper::vgetlow(data_q16))); - vres2 = wrapper::vadd(vres2, wrapper::vmovl(wrapper::vgethigh(data_q16))); - vres3 = wrapper::vadd(vres3, wrapper::vmovl(wrapper::vgetlow(data2_q16))); - vres4 = wrapper::vadd(vres4, wrapper::vmovl(wrapper::vgethigh(data2_q16))); - } - } - - if(input_qinfo != output_qinfo) - { - const float32x4x4_t vres = - { - { - vcvtq_f32_q32(vres1), - vcvtq_f32_q32(vres2), - vcvtq_f32_q32(vres3), - vcvtq_f32_q32(vres4), - } - }; - const auto requantized_output = vrequantize_pooling_with_scale(vres, quant_rescale, scale, new_offset); - // Store result - wrapper::vstore(reinterpret_cast(output.ptr()) + x_off, wrapper::vgetlow(requantized_output)); - wrapper::vstore(reinterpret_cast(output.ptr()) + x_off + 8, wrapper::vgethigh(requantized_output)); - } - else - { - const float32x4_t scale_v = vdupq_n_f32(scale); - // Divide by scale and add 0.5f to round to nearest instead of rounding towards zero - vres1 = vcvtq_q32_f32(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres1), scale_v)); - vres2 = vcvtq_q32_f32(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres2), scale_v)); - vres3 = vcvtq_q32_f32(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres3), scale_v)); - vres4 = vcvtq_q32_f32(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres4), scale_v)); - - const q8x8_t res1 = wrapper::vmovn(wrapper::vcombine(wrapper::vmovn(vres1), wrapper::vmovn(vres2))); - const q8x8_t res2 = wrapper::vmovn(wrapper::vcombine(wrapper::vmovn(vres3), wrapper::vmovn(vres4))); - // Store result - wrapper::vstore(reinterpret_cast(output.ptr()) + x_off, res1); - wrapper::vstore(reinterpret_cast(output.ptr()) + x_off + 8, res2); - } - } - else - { - q8x16_t vres = wrapper::vdup_n(std::numeric_limits::min(), wrapper::traits::vector_128_tag{}); - - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const q8x16_t data = wrapper::vloadq(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z())) + x_off); - vres = wrapper::vmax(vres, data); - } - } - - // Store result - wrapper::vstore(reinterpret_cast(output.ptr()) + x_off, (input_qinfo != output_qinfo) ? vrequantize_pooling(wrapper::vgetlow(vres), wrapper::vgethigh(vres), - requant_qinfo) : - vres); - } - } - - if(pooling_type == PoolingType::MAX) - { - for(; x_off <= (window_end_x - window_half_step_x); x_off += window_half_step_x) - { - q8x8_t vres = wrapper::vdup_n(std::numeric_limits::min(), wrapper::traits::vector_64_tag{}); - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const q8x8_t data = wrapper::vload(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z())) + x_off); - vres = wrapper::vmax(vres, data); - } - } - - // Store result - wrapper::vstore(reinterpret_cast(output.ptr()) + x_off, - (input_qinfo != output_qinfo) ? vrequantize_pooling(vres, requant_qinfo) : vres); - } - } - - // Left-overs loop - for(; x_off < window_end_x; ++x_off) - { - if(pooling_type != PoolingType::MAX) - { - q32_t res = static_cast(0.f); - - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, - pool_stride_y); - - // Perform pooling - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const T data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z())) + x_off); - res += data; - } - } - - if(input_qinfo != output_qinfo) - { - const float res_f = static_cast(res); - const float new_scale = quant_rescale / scale; - const auto requantized_output = quantize(res_f, UniformQuantizationInfo(new_scale, new_offset)); - - // Store result - *(reinterpret_cast(output.ptr()) + x_off) = requantized_output; - } - else - { - // Divide by scale and add 0.5f to round to nearest instead of rounding towards zero - res = static_cast(0.5f + static_cast(res) * scale); - - // Store result - *(reinterpret_cast(output.ptr()) + x_off) = res; - } - } - else - { - T res = std::numeric_limits::min(); - - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const T data = *(reinterpret_cast(input.ptr() + (x - pool_pad_left) * static_cast(_input->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast - (_input->info()->strides_in_bytes().z())) + x_off); - res = std::max(res, data); - } - } - - // Store result - if(input_qinfo != output_qinfo) - { - const float res_f = static_cast(res); - *(reinterpret_cast(output.ptr()) + x_off) = quantize(res_f, requant_qinfo); - } - else - { - *(reinterpret_cast(output.ptr()) + x_off) = res; - } - } - } - - }, - input, output); -} - -Status NEPoolingLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, const ITensorInfo *indices) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); - - unsigned int pooled_w = 0; - unsigned int pooled_h = 0; - unsigned int num_elems_processed_per_iteration = 0; - BorderSize border_size(0); - - const bool is_global_pooling = pool_info.is_global_pooling; - unsigned int pool_size_x = 0; - unsigned int pool_size_y = 0; - - // Get data layout - const auto data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? input->data_layout() : pool_info.data_layout; - const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); - - pool_size_x = is_global_pooling ? input->dimension(idx_width) : pool_info.pool_size.width; - pool_size_y = is_global_pooling ? input->dimension(idx_height) : pool_info.pool_size.height; - - // Validate pool info before calling scaled_dimensions - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_pool_info(pool_size_x, pool_size_y)); - - // Check output dimensions - std::tie(pooled_w, pooled_h) = scaled_dimensions(input->dimension(idx_width), - input->dimension(idx_height), - pool_size_x, - pool_size_y, - pool_info.pad_stride_info); - - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, pool_info, pooled_w, pooled_h, indices, Size2D(pool_size_x, pool_size_y))); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), - (indices) ? indices->clone().get() : nullptr, pool_info, num_elems_processed_per_iteration, border_size, pooled_w, pooled_h, - pool_size_x, pool_size_y) - .first); - - return Status{}; -} - -void NEPoolingLayerKernel::run(const Window &window, const ThreadInfo &info) -{ - ARM_COMPUTE_UNUSED(info); - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); - ARM_COMPUTE_ERROR_ON(_func == nullptr); - - const unsigned int pool_stride_x = _pool_info.pad_stride_info.stride().first; - const unsigned int pool_stride_y = _pool_info.pad_stride_info.stride().second; - const unsigned int pool_size = _pool_info.pool_size.width; - const bool exclude_padding = _pool_info.exclude_padding; - - Window window_input(window); - if(_data_layout == DataLayout::NCHW) - { - // Set step for input in x and y direction for the input - unsigned int window_x_inc = 0; - switch(_input->info()->data_type()) - { - case DataType::QASYMM8: - case DataType::QASYMM8_SIGNED: - { - window_x_inc = pool_stride_x; - if((pool_size == 2 || pool_size == 3) && pool_stride_x < 3) - { - window_x_inc = (pool_stride_x == 2) ? _num_elems_processed_per_iteration * 2 : _num_elems_processed_per_iteration; - } - break; - } - - case DataType::F16: - case DataType::F32: - { - window_x_inc = pool_stride_x; - break; - } - default: - { - ARM_COMPUTE_ERROR("Not supported"); - } - } - window_input.set(Window::DimX, Window::Dimension(window.x().start() * pool_stride_x, window.x().end() * pool_stride_x, window_x_inc)); - window_input.set(Window::DimY, Window::Dimension(window.y().start() * pool_stride_y, window.y().end() * pool_stride_y, pool_stride_y)); - } - else - { - window_input.set(Window::DimX, Window::Dimension(0, 1, 1)); - window_input.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(1), pool_stride_x)); - window_input.set(Window::DimZ, Window::Dimension(0, _input->info()->dimension(2), pool_stride_y)); - } - - // Run function - (this->*_func)(window_input, window, _pool_info.pool_type, exclude_padding); -} -} // namespace arm_compute diff --git a/src/core/NEON/kernels/NEPoolingLayerKernel.h b/src/core/NEON/kernels/NEPoolingLayerKernel.h deleted file mode 100644 index aa3d2f3f01..0000000000 --- a/src/core/NEON/kernels/NEPoolingLayerKernel.h +++ /dev/null @@ -1,229 +0,0 @@ -/* - * Copyright (c) 2017-2020 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_NEPOOLINGLAYERKERNEL_H -#define ARM_COMPUTE_NEPOOLINGLAYERKERNEL_H - -#include "src/core/NEON/INEKernel.h" - -namespace arm_compute -{ -class ITensor; - -/** Interface for the pooling layer kernel */ -class NEPoolingLayerKernel : public INEKernel -{ -public: - const char *name() const override - { - return "NEPoolingLayerKernel"; - } - /** Default constructor */ - NEPoolingLayerKernel(); - /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEPoolingLayerKernel(const NEPoolingLayerKernel &) = delete; - /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEPoolingLayerKernel &operator=(const NEPoolingLayerKernel &) = delete; - /** Allow instances of this class to be moved */ - NEPoolingLayerKernel(NEPoolingLayerKernel &&) = default; - /** Allow instances of this class to be moved */ - NEPoolingLayerKernel &operator=(NEPoolingLayerKernel &&) = default; - /** Default destructor */ - ~NEPoolingLayerKernel() = default; - /** Set the input and output tensors. - * - * @note F16 are supported for pool sizes 2 and 3 only - * - * @param[in] input Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. - * @param[out] output Destination tensor. Data types supported: Same as @p input. - * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo. - * @param[out] indices (optional) The indices of the maximal values. Data type supported: U32. - */ - void configure(const ITensor *input, ITensor *output, const PoolingLayerInfo &pool_info, ITensor *indices = nullptr); - /** Static function to check if given info will lead to a valid configuration of @ref NEPoolingLayerKernel - * - * @note F16 are supported for pool sizes 2 and 3 only - * - * @param[in] input Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. - * @param[in] output Destination tensor. Data types supported: Same as @p input. - * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo. - * @param[in] indices (optional) The indices of the maximal values. Data type supported: U32. - * - * @return a status - */ - static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, const ITensorInfo *indices = nullptr); - - // Inherited methods overridden: - void run(const Window &window, const ThreadInfo &info) override; - BorderSize border_size() const override; - -private: - /** Function to perform 2x2 pooling. - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - * @param[in] pooling_type Pooling operation to be computed. - * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. - */ - void pooling2_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); - /** Function to perform 2x2 pooling and compute the pooling indices. The indices can be used for max unpool. - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - */ - void pooling2_f32_nhwc_maxpool_indices(const Window &window_input, const Window &window); - /** Function to perform MxN pooling for 32-bit floating point values. - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - * @param[in] pooling_type Pooling operation to be computed. - * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. - */ - void poolingMxN_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); - /** Function to perform MxN pooling for 32-bit floating point values (NHWC). - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - * @param[in] pooling_type Pooling operation to be computed. - * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. - */ - void poolingMxN_f32_nhwc(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); - /** Function to perform 7x7 pooling. - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - * @param[in] pooling_type Pooling operation to be computed. - * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. - */ - void pooling7_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); - /** Function to perform 3x3 pooling. - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - * @param[in] pooling_type Pooling operation to be computed. - * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. - */ - void pooling3_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); - /** Function to perform 2x2 pooling for float16_t. - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - * @param[in] pooling_type Pooling operation to be computed. - * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. - */ - void pooling2_f16_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); - /** Function to perform 2x2 pooling and compute the pooling indices for FP32/FP16. The indices can be used for max unpool. - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - */ - template - void pooling2_nchw_maxpool_indices(const Window &window_input, const Window &window); - /** Function to perform 2x2 pooling and compute the pooling indices. The indices can be used for max unpool. - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - */ - void pooling2_f16_nhwc_maxpool_indices(const Window &window_input, const Window &window); - /** Function to perform 3x3 pooling. - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - * @param[in] pooling_type Pooling operation to be computed. - * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. - */ - void pooling3_f16_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); - /** Function to perform MxN pooling for 16-bit floating point values. - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - * @param[in] pooling_type Pooling operation to be computed. - * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. - */ - void poolingMxN_f16_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); - /** Function to perform MxN pooling for 16-bit floating point values. (NHWC) - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - * @param[in] pooling_type Pooling operation to be computed. - * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. - */ - void poolingMxN_f16_nhwc(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); - /** Template function to perform 2x2 pooling for 8bit quantized fixed point. (NCHW) - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - * @param[in] pooling_type Pooling operation to be computed. - * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. - */ - template - void pooling2_q8_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); - /** Template function to perform 3x3 pooling for 8bit quantized fixed point. (NCHW) - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - * @param[in] pooling_type Pooling operation to be computed. - * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. - */ - template - void pooling3_q8_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); - /** Template function to perform MxN pooling for 8-bit quantized. (NCHW) - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - * @param[in] pooling_type Pooling operation to be computed. - * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. - */ - template - void poolingMxN_q8_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); - /** Template function to perform MxN pooling for 8-bit quantized. (NHWC) - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - * @param[in] pooling_type Pooling operation to be computed. - * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. - */ - template - void poolingMxN_q8_nhwc(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); - /** Common signature for all the specialised Pooling functions - * - * @param[in] window_input Input region on which to execute the kernel. - * @param[in] window Output region on which to execute the kernel. - * @param[in] pooling_type Pooling operation to be computed. - * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. - */ - using PoolingFunction = void (NEPoolingLayerKernel::*)(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding); - -private: - PoolingFunction _func; - const ITensor *_input; - ITensor *_output; - ITensor *_indices; - PoolingLayerInfo _pool_info; - DataLayout _data_layout; - unsigned int _num_elems_processed_per_iteration; - BorderSize _border_size; - bool _is_square; -}; -} // namespace arm_compute -#endif /*ARM_COMPUTE_NEPOOLINGLAYERKERNEL_H */ diff --git a/src/core/NEON/kernels/assembly/NEPoolingAssemblyWrapperKernel.cpp b/src/core/NEON/kernels/assembly/NEPoolingAssemblyWrapperKernel.cpp deleted file mode 100644 index 04406663fc..0000000000 --- a/src/core/NEON/kernels/assembly/NEPoolingAssemblyWrapperKernel.cpp +++ /dev/null @@ -1,269 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/NEON/kernels/assembly/NEPoolingAssemblyWrapperKernel.h" -#include "arm_compute/core/Utils.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "arm_compute/core/utils/quantization/AsymmHelpers.h" -#include "src/core/CPP/Validate.h" -#include "src/core/helpers/AutoConfiguration.h" -#include "src/core/helpers/WindowHelpers.h" - -#include - -namespace arm_compute -{ -using namespace arm_compute::misc::shape_calculator; - -void NEPoolingAssemblyWrapperKernel::configure(const ITensorInfo *input, ITensorInfo *output, const PoolingLayerInfo &info, const CPUInfo &cpu_info) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - - // Output initialization if not yet initialized - auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_pool_shape(*input, info))); - - const bool requantize = input->quantization_info() != output->quantization_info(); - - switch(input->data_type()) - { - case DataType::QASYMM8: - if(requantize) - { - create_arm_pooling_requant(input, output, info, cpu_info); - } - else - { - create_arm_pooling(input, output, info, cpu_info); - } - break; - case DataType::QASYMM8_SIGNED: - if(requantize) - { - create_arm_pooling_requant(input, output, info, cpu_info); - } - else - { - create_arm_pooling(input, output, info, cpu_info); - } - break; -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - case DataType::F16: - create_arm_pooling(input, output, info, cpu_info); - break; -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - case DataType::F32: - create_arm_pooling(input, output, info, cpu_info); - break; - default: - break; - } - - Window win = calculate_max_window(*output, Steps()); - INEKernel::configure(win); -} - -Status NEPoolingAssemblyWrapperKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); - -#ifndef __aarch64__ - ARM_COMPUTE_RETURN_ERROR_MSG("32-bit is not supported by assembly kernels"); -#endif /* __aarch64__ */ - ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((input->data_layout() != DataLayout::NHWC) || (info.data_layout != DataLayout::NHWC), "Only NHWC is supported by assembly kernels"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((info.pool_type != PoolingType::AVG) && (info.pool_type != PoolingType::MAX), - "Only AVG and MAX pooling are supported by assembly kernels"); - - if(output->total_size() > 0) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - - const auto input_qinfo = input->quantization_info().uniform(); - const auto output_qinfo = output->quantization_info().uniform(); - - if(input_qinfo != output_qinfo) - { - const float multiplier = input_qinfo.scale / output_qinfo.scale; - int32_t output_multiplier{}; - int32_t output_shift{}; - ARM_COMPUTE_RETURN_ERROR_ON(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift)); - } - else - { - if(input->data_type() == DataType::QASYMM8) - { - const bool has_padding = info.pad_stride_info.has_padding(); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(!info.exclude_padding && has_padding, "Assembly kernels do not support padding for QASYMM8 with same input/output quantization info"); - } - } - } - else - { - if(input->data_type() == DataType::QASYMM8) - { - // If output is not configured, the quantization info are the same - const bool has_padding = info.pad_stride_info.has_padding(); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(!info.exclude_padding && has_padding, "Assembly kernels do not support padding for QASYMM8 with same input/output quantization info"); - } - } - return Status{}; -} - -void NEPoolingAssemblyWrapperKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(_kernel_asm.get()); - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_UNUSED(window); - ARM_COMPUTE_UNUSED(info); - - ARM_COMPUTE_ERROR_ON(tensors.empty()); - - const ITensor *input = tensors.get_const_tensor(TensorType::ACL_SRC); - ITensor *output = tensors.get_tensor(TensorType::ACL_DST_0); - ITensor *workspace = tensors.get_tensor(TensorType::ACL_DST_1); - - const auto in_ptr = input->buffer() + input->info()->offset_first_element_in_bytes(); - auto out_ptr = output->buffer() + output->info()->offset_first_element_in_bytes(); - auto working_space = workspace->buffer() + workspace->info()->offset_first_element_in_bytes(); - - const auto input_shape = input->info()->tensor_shape(); - const auto output_shape = output->info()->tensor_shape(); - const auto input_padding = input->info()->padding(); - const auto output_padding = output->info()->padding(); - - const size_t ld_input_col = input_shape[0] + input_padding.left + input_padding.right; - const size_t ld_input_row = ld_input_col * (input_shape[1] + input_padding.top + input_padding.bottom); - const size_t ld_input_batch = ld_input_row * input_shape[2]; - const size_t ld_output_col = output_shape[0] + output_padding.right; - const size_t ld_output_row = ld_output_col * (output_shape[1] + output_padding.top + output_padding.bottom); - const size_t ld_output_batch = ld_output_row * output_shape[2]; - - _kernel_asm->execute(in_ptr, ld_input_col, ld_input_row, ld_input_batch, - out_ptr, ld_output_col, ld_output_row, ld_output_batch, - working_space, info.thread_id, info.num_threads); -} - -size_t NEPoolingAssemblyWrapperKernel::get_working_size(unsigned int num_threads) const -{ - return _kernel_asm->get_working_size(num_threads); -} - -bool NEPoolingAssemblyWrapperKernel::is_configured() const -{ - return _kernel_asm != nullptr; -} - -template -void NEPoolingAssemblyWrapperKernel::create_arm_pooling(const ITensorInfo *input, ITensorInfo *output, const PoolingLayerInfo &info, const CPUInfo &cpu_info) -{ - const arm_conv::pooling::PoolingType pool_type = (info.pool_type == PoolingType::AVG) ? arm_conv::pooling::PoolingType::AVERAGE : arm_conv::pooling::PoolingType::MAX; - - arm_conv::pooling::PoolingWindow window{}; - window.cols = static_cast(info.pool_size.x()); - window.rows = static_cast(info.pool_size.y()); - - arm_conv::pooling::PoolingStride stride{}; - std::tie(stride.cols, stride.rows) = info.pad_stride_info.stride(); - - const arm_conv::pooling::PaddingValues padding{ info.pad_stride_info.pad_left(), info.pad_stride_info.pad_top(), info.pad_stride_info.pad_right(), info.pad_stride_info.pad_bottom() }; - - constexpr unsigned int idx_width = 1; - constexpr unsigned int idx_height = 2; - constexpr unsigned int idx_channels = 0; - constexpr unsigned int idx_batches = 3; - - const unsigned int n_batches = input->dimension(idx_batches); - const unsigned int input_rows = input->dimension(idx_height); - const unsigned int input_cols = input->dimension(idx_width); - const unsigned int n_channels = input->dimension(idx_channels); - const unsigned int output_rows = output->dimension(idx_height); - const unsigned int output_cols = output->dimension(idx_width); - - arm_conv::pooling::PoolingArgs args(&cpu_info, pool_type, window, stride, info.exclude_padding, n_batches, input_rows, input_cols, n_channels, output_rows, output_cols, padding, nullptr); - - // Configure assembly pooling kernel - auto pooling_kernel_asm = arm_conv::pooling::pooling(args); - if(pooling_kernel_asm == nullptr) - { - // Configuration not supported: Leave function unconfigured: - return; - } - - _kernel_asm = std::move(pooling_kernel_asm); -} - -template -void NEPoolingAssemblyWrapperKernel::create_arm_pooling_requant(const ITensorInfo *input, ITensorInfo *output, const PoolingLayerInfo &info, const CPUInfo &cpu_info) -{ - const arm_conv::pooling::PoolingType pool_type = (info.pool_type == PoolingType::AVG) ? arm_conv::pooling::PoolingType::AVERAGE : arm_conv::pooling::PoolingType::MAX; - - arm_conv::pooling::PoolingWindow window{}; - window.cols = static_cast(info.pool_size.x()); - window.rows = static_cast(info.pool_size.y()); - - arm_conv::pooling::PoolingStride stride{}; - std::tie(stride.cols, stride.rows) = info.pad_stride_info.stride(); - - const arm_conv::pooling::PaddingValues padding{ info.pad_stride_info.pad_left(), info.pad_stride_info.pad_top(), info.pad_stride_info.pad_right(), info.pad_stride_info.pad_bottom() }; - - constexpr unsigned int idx_width = 1; - constexpr unsigned int idx_height = 2; - constexpr unsigned int idx_channels = 0; - constexpr unsigned int idx_batches = 3; - - const unsigned int n_batches = input->dimension(idx_batches); - const unsigned int input_rows = input->dimension(idx_height); - const unsigned int input_cols = input->dimension(idx_width); - const unsigned int n_channels = input->dimension(idx_channels); - const unsigned int output_rows = output->dimension(idx_height); - const unsigned int output_cols = output->dimension(idx_width); - - arm_conv::pooling::PoolingArgs args(&cpu_info, pool_type, window, stride, info.exclude_padding, n_batches, input_rows, input_cols, n_channels, output_rows, output_cols, padding, nullptr); - - const auto input_qinfo = input->quantization_info().uniform(); - const auto output_qinfo = output->quantization_info().uniform(); - - const float multiplier = input_qinfo.scale / output_qinfo.scale; - int32_t output_multiplier{}; - int32_t output_shift{}; - quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift); - - const arm_conv::pooling::Requantize32 requant_args(input_qinfo.offset, - output_qinfo.offset, - output_shift, // left shift - 0, // right shift - output_multiplier); - - // Configure assembly pooling kernel with requantization - auto pooling_kernel_asm = arm_conv::pooling::pooling(args, requant_args); - if(pooling_kernel_asm == nullptr) - { - // Configuration not supported: Leave function unconfigured: - return; - } - - _kernel_asm = std::move(pooling_kernel_asm); -} -} // namespace arm_compute diff --git a/src/core/NEON/kernels/assembly/NEPoolingAssemblyWrapperKernel.h b/src/core/NEON/kernels/assembly/NEPoolingAssemblyWrapperKernel.h deleted file mode 100644 index b2fa5b5714..0000000000 --- a/src/core/NEON/kernels/assembly/NEPoolingAssemblyWrapperKernel.h +++ /dev/null @@ -1,116 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_ASSEMBLY_POOLING_KERNEL_WRAPPER_KERNEL_H -#define ARM_COMPUTE_ASSEMBLY_POOLING_KERNEL_WRAPPER_KERNEL_H - -#include "src/core/NEON/INEKernel.h" -#include "src/core/NEON/kernels/assembly/pooling.hpp" - -#include "pool_common.hpp" - -namespace arm_compute -{ -class ITensor; - -/** This class is a wrapper for the assembly kernels. - * - * Some kernels were written in assembly and highly optimised for specific - * CPUs like A53 or A55. The arm compute library creates an instance of - * NEPoolingAssemblyWrapperKernel and other auxiliary data structures to - * execute a single assembly kernel in the context of an NEFunction. - * - */ -class NEPoolingAssemblyWrapperKernel final : public INEKernel -{ -public: - /** Constructor - */ - NEPoolingAssemblyWrapperKernel() = default; - NEPoolingAssemblyWrapperKernel(NEPoolingAssemblyWrapperKernel &) = delete; - NEPoolingAssemblyWrapperKernel(NEPoolingAssemblyWrapperKernel &&) = default; - NEPoolingAssemblyWrapperKernel &operator=(NEPoolingAssemblyWrapperKernel &) = delete; - - const char *name() const override - { - return "NEPoolingAssemblyWrapperKernel"; - } - - /** Initialise the kernel's input and output. - * - * @param[in] input Input tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. - * @param[out] output Output tensor to store the result of pooling. Data types supported: same as @p input. - * @param[in] info Pooling meta-data - */ - void configure(const ITensorInfo *input, ITensorInfo *output, const PoolingLayerInfo &info, const CPUInfo &cpu_info); - - /** Indicates whether or not this function can be used to process the given parameters. - * - * @param[in] input Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. - * @param[in] output Output tensor to store the result of pooling. Data types supported: same as @p input. - * @param[in] info Pooling meta-data - * - * @return a status. - */ - static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &info); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override; - - /** Get size of the workspace needed by the assembly kernel. - * - * @param[in] num_threads Maximum number of threads that are going to be spawned. - * - * @return size of workspace - */ - size_t get_working_size(unsigned int num_threads) const; - - /** Was the asm kernel successfully configured? - * - * @return True if the asm kernel is configured and ready to run - */ - bool is_configured() const; - -private: - /** Helper function to create the assembly kernel. - * - * @param[in] input Input tensor info. - * @param[in] output Output tensor info. - * @param[in] info Pooling layer meta-data. - */ - template - void create_arm_pooling(const ITensorInfo *input, ITensorInfo *output, const PoolingLayerInfo &info, const CPUInfo &cpu_info); - - /** Helper function to create the assembly kernel with requantization support - * - * @param[in] input Input tensor info. - * @param[in] output Output tensor info. - * @param[in] info Pooling layer meta-data. - */ - template - void create_arm_pooling_requant(const ITensorInfo *input, ITensorInfo *output, const PoolingLayerInfo &info, const CPUInfo &cpu_info); - - std::unique_ptr _kernel_asm{ nullptr }; -}; -} // namespace arm_compute -#endif /* ARM_COMPUTE_ASSEMBLY_POOLING_KERNEL_WRAPPER_KERNEL_H */ diff --git a/src/core/cpu/kernels/CpuPoolingAssemblyWrapperKernel.cpp b/src/core/cpu/kernels/CpuPoolingAssemblyWrapperKernel.cpp new file mode 100644 index 0000000000..19a0e90d0e --- /dev/null +++ b/src/core/cpu/kernels/CpuPoolingAssemblyWrapperKernel.cpp @@ -0,0 +1,276 @@ +/* + * Copyright (c) 2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "src/core/cpu/kernels/CpuPoolingAssemblyWrapperKernel.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/core/utils/quantization/AsymmHelpers.h" +#include "src/core/CPP/Validate.h" +#include "src/core/NEON/INEKernel.h" +#include "src/core/helpers/AutoConfiguration.h" +#include "src/core/helpers/WindowHelpers.h" + +#include + +namespace arm_compute +{ +namespace cpu +{ +namespace kernels +{ +using namespace arm_compute::misc::shape_calculator; + +void CpuPoolingAssemblyWrapperKernel::configure(const ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &info, const CPUInfo &cpu_info) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); + + // dst initialization if not yet initialized + auto_init_if_empty(*dst, src->clone()->set_tensor_shape(compute_pool_shape(*src, info))); + + const bool requantize = src->quantization_info() != dst->quantization_info(); + + switch(src->data_type()) + { + case DataType::QASYMM8: + if(requantize) + { + create_arm_pooling_requant(src, dst, info, cpu_info); + } + else + { + create_arm_pooling(src, dst, info, cpu_info); + } + break; + case DataType::QASYMM8_SIGNED: + if(requantize) + { + create_arm_pooling_requant(src, dst, info, cpu_info); + } + else + { + create_arm_pooling(src, dst, info, cpu_info); + } + break; +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + case DataType::F16: + create_arm_pooling(src, dst, info, cpu_info); + break; +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + case DataType::F32: + create_arm_pooling(src, dst, info, cpu_info); + break; + default: + break; + } + + Window win = calculate_max_window(*dst, Steps()); + INEKernel::configure(win); +} + +Status CpuPoolingAssemblyWrapperKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); + +#ifndef __aarch64__ + ARM_COMPUTE_RETURN_ERROR_MSG("32-bit is not supported by assembly kernels"); +#endif /* __aarch64__ */ + ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((src->data_layout() != DataLayout::NHWC) || (info.data_layout != DataLayout::NHWC), "Only NHWC is supported by assembly kernels"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((info.pool_type != PoolingType::AVG) && (info.pool_type != PoolingType::MAX), + "Only AVG and MAX pooling are supported by assembly kernels"); + + if(dst->total_size() > 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); + + const auto src_qinfo = src->quantization_info().uniform(); + const auto dst_qinfo = dst->quantization_info().uniform(); + + if(src_qinfo != dst_qinfo) + { + const float multiplier = src_qinfo.scale / dst_qinfo.scale; + int32_t dst_multiplier{}; + int32_t dst_shift{}; + ARM_COMPUTE_RETURN_ERROR_ON(quantization::calculate_quantized_multiplier(multiplier, &dst_multiplier, &dst_shift)); + } + else + { + if(src->data_type() == DataType::QASYMM8) + { + const bool has_padding = info.pad_stride_info.has_padding(); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(!info.exclude_padding && has_padding, "Assembly kernels do not support padding for QASYMM8 with same src/dst quantization info"); + } + } + } + else + { + if(src->data_type() == DataType::QASYMM8) + { + // If dst is not configured, the quantization info are the same + const bool has_padding = info.pad_stride_info.has_padding(); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(!info.exclude_padding && has_padding, "Assembly kernels do not support padding for QASYMM8 with same src/dst quantization info"); + } + } + return Status{}; +} + +void CpuPoolingAssemblyWrapperKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(_kernel_asm.get()); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_UNUSED(window); + ARM_COMPUTE_UNUSED(info); + + ARM_COMPUTE_ERROR_ON(tensors.empty()); + + const ITensor *src = tensors.get_const_tensor(TensorType::ACL_SRC); + ITensor *dst = tensors.get_tensor(TensorType::ACL_DST_0); + ITensor *workspace = tensors.get_tensor(TensorType::ACL_DST_1); + + const auto in_ptr = src->buffer() + src->info()->offset_first_element_in_bytes(); + auto out_ptr = dst->buffer() + dst->info()->offset_first_element_in_bytes(); + auto working_space = workspace->buffer() + workspace->info()->offset_first_element_in_bytes(); + + const auto src_shape = src->info()->tensor_shape(); + const auto dst_shape = dst->info()->tensor_shape(); + const auto src_padding = src->info()->padding(); + const auto dst_padding = dst->info()->padding(); + + const size_t ld_src_col = src_shape[0] + src_padding.left + src_padding.right; + const size_t ld_src_row = ld_src_col * (src_shape[1] + src_padding.top + src_padding.bottom); + const size_t ld_src_batch = ld_src_row * src_shape[2]; + const size_t ld_dst_col = dst_shape[0] + dst_padding.left + dst_padding.right; + const size_t ld_dst_row = ld_dst_col * (dst_shape[1] + dst_padding.top + dst_padding.bottom); + const size_t ld_dst_batch = ld_dst_row * dst_shape[2]; + + _kernel_asm->execute(in_ptr, ld_src_col, ld_src_row, ld_src_batch, + out_ptr, ld_dst_col, ld_dst_row, ld_dst_batch, + working_space, info.thread_id, info.num_threads); +} + +size_t CpuPoolingAssemblyWrapperKernel::get_working_size(unsigned int num_threads) const +{ + return _kernel_asm->get_working_size(num_threads); +} + +bool CpuPoolingAssemblyWrapperKernel::is_configured() const +{ + return _kernel_asm != nullptr; +} + +template +void CpuPoolingAssemblyWrapperKernel::create_arm_pooling(const ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &info, const CPUInfo &cpu_info) +{ + const arm_conv::pooling::PoolingType pool_type = (info.pool_type == PoolingType::AVG) ? arm_conv::pooling::PoolingType::AVERAGE : arm_conv::pooling::PoolingType::MAX; + + arm_conv::pooling::PoolingWindow window{}; + window.cols = static_cast(info.pool_size.x()); + window.rows = static_cast(info.pool_size.y()); + + arm_conv::pooling::PoolingStride stride{}; + std::tie(stride.cols, stride.rows) = info.pad_stride_info.stride(); + + const arm_conv::pooling::PaddingValues padding{ info.pad_stride_info.pad_left(), info.pad_stride_info.pad_top(), info.pad_stride_info.pad_right(), info.pad_stride_info.pad_bottom() }; + + constexpr unsigned int idx_width = 1; + constexpr unsigned int idx_height = 2; + constexpr unsigned int idx_channels = 0; + constexpr unsigned int idx_batches = 3; + + const unsigned int n_batches = src->dimension(idx_batches); + const unsigned int src_rows = src->dimension(idx_height); + const unsigned int src_cols = src->dimension(idx_width); + const unsigned int n_channels = src->dimension(idx_channels); + const unsigned int dst_rows = dst->dimension(idx_height); + const unsigned int dst_cols = dst->dimension(idx_width); + + arm_conv::pooling::PoolingArgs args(&cpu_info, pool_type, window, stride, info.exclude_padding, n_batches, src_rows, src_cols, n_channels, dst_rows, dst_cols, padding, nullptr); + + // Configure assembly pooling kernel + auto pooling_kernel_asm = arm_conv::pooling::pooling(args); + if(pooling_kernel_asm == nullptr) + { + // Configuration not supported: Leave function unconfigured: + return; + } + + _kernel_asm = std::move(pooling_kernel_asm); +} + +template +void CpuPoolingAssemblyWrapperKernel::create_arm_pooling_requant(const ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &info, const CPUInfo &cpu_info) +{ + const arm_conv::pooling::PoolingType pool_type = (info.pool_type == PoolingType::AVG) ? arm_conv::pooling::PoolingType::AVERAGE : arm_conv::pooling::PoolingType::MAX; + + arm_conv::pooling::PoolingWindow window{}; + window.cols = static_cast(info.pool_size.x()); + window.rows = static_cast(info.pool_size.y()); + + arm_conv::pooling::PoolingStride stride{}; + std::tie(stride.cols, stride.rows) = info.pad_stride_info.stride(); + + const arm_conv::pooling::PaddingValues padding{ info.pad_stride_info.pad_left(), info.pad_stride_info.pad_top(), info.pad_stride_info.pad_right(), info.pad_stride_info.pad_bottom() }; + + constexpr unsigned int idx_width = 1; + constexpr unsigned int idx_height = 2; + constexpr unsigned int idx_channels = 0; + constexpr unsigned int idx_batches = 3; + + const unsigned int n_batches = src->dimension(idx_batches); + const unsigned int src_rows = src->dimension(idx_height); + const unsigned int src_cols = src->dimension(idx_width); + const unsigned int n_channels = src->dimension(idx_channels); + const unsigned int dst_rows = dst->dimension(idx_height); + const unsigned int dst_cols = dst->dimension(idx_width); + + arm_conv::pooling::PoolingArgs args(&cpu_info, pool_type, window, stride, info.exclude_padding, n_batches, src_rows, src_cols, n_channels, dst_rows, dst_cols, padding, nullptr); + + const auto src_qinfo = src->quantization_info().uniform(); + const auto dst_qinfo = dst->quantization_info().uniform(); + + const float multiplier = src_qinfo.scale / dst_qinfo.scale; + int32_t dst_multiplier{}; + int32_t dst_shift{}; + quantization::calculate_quantized_multiplier(multiplier, &dst_multiplier, &dst_shift); + + const arm_conv::pooling::Requantize32 requant_args(src_qinfo.offset, + dst_qinfo.offset, + dst_shift, // left shift + 0, // right shift + dst_multiplier); + + // Configure assembly pooling kernel with requantization + auto pooling_kernel_asm = arm_conv::pooling::pooling(args, requant_args); + if(pooling_kernel_asm == nullptr) + { + // Configuration not supported: Leave function unconfigured: + return; + } + + _kernel_asm = std::move(pooling_kernel_asm); +} +} // namespace kernels +} // namespace cpu +} // namespace arm_compute diff --git a/src/core/cpu/kernels/CpuPoolingAssemblyWrapperKernel.h b/src/core/cpu/kernels/CpuPoolingAssemblyWrapperKernel.h new file mode 100644 index 0000000000..34ec452deb --- /dev/null +++ b/src/core/cpu/kernels/CpuPoolingAssemblyWrapperKernel.h @@ -0,0 +1,123 @@ +/* + * Copyright (c) 2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ARM_COMPUTE_CPU_POOLING_ASSEMBLY_WRAPPER_KERNEL_H +#define ARM_COMPUTE_CPU_POOLING_ASSEMBLY_WRAPPER_KERNEL_H + +#include "arm_compute/core/Types.h" +#include "src/core/NEON/kernels/assembly/pooling.hpp" +#include "src/core/common/Macros.h" +#include "src/core/cpu/ICpuKernel.h" + +#include "pool_common.hpp" + +namespace arm_compute +{ +namespace cpu +{ +namespace kernels +{ +/** This class is a wrapper for the assembly kernels. + * + * Some kernels were written in assembly and highly optimised for specific + * CPUs like A53 or A55. The arm compute library creates an instance of + * CpuPoolingAssemblyWrapperKernel and other auxiliary data structures to + * execute a single assembly kernel in the context of an NEFunction. + * + */ +class CpuPoolingAssemblyWrapperKernel final : public ICpuKernel +{ +public: + /** Constructor + */ + CpuPoolingAssemblyWrapperKernel() = default; + CpuPoolingAssemblyWrapperKernel(CpuPoolingAssemblyWrapperKernel &) = delete; + CpuPoolingAssemblyWrapperKernel(CpuPoolingAssemblyWrapperKernel &&) = default; + CpuPoolingAssemblyWrapperKernel &operator=(CpuPoolingAssemblyWrapperKernel &) = delete; + + const char *name() const override + { + return "CpuPoolingAssemblyWrapperKernel"; + } + + /** Initialise the kernel's src and dst. + * + * @param[in] src Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. + * @param[out] dst Destination tensor info to store the result of pooling. Data types supported: same as @p src. + * @param[in] info Pooling meta-data. + * @param[in] cpu_info CPU information needed to select the most appropriate kernel. + */ + void configure(const ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &info, const CPUInfo &cpu_info); + + /** Indicates whether or not this function can be used to process the given parameters. + * + * @param[in] src Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. + * @param[in] dst Destination tensor to store the result of pooling. Data types supported: same as @p src. + * @param[in] info Pooling meta-data + * + * @return a status. + */ + static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &info); + + // Inherited methods overridden: + void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override; + + /** Get size of the workspace needed by the assembly kernel. + * + * @param[in] num_threads Maximum number of threads that are going to be spawned. + * + * @return size of workspace + */ + size_t get_working_size(unsigned int num_threads) const; + + /** Was the asm kernel successfully configured? + * + * @return True if the asm kernel is configured and ready to run + */ + bool is_configured() const; + +private: + /** Helper function to create the assembly kernel. + * + * @param[in] src Source tensor info. + * @param[in] dst Destination tensor info. + * @param[in] info Pooling layer meta-data. + */ + template + void create_arm_pooling(const ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &info, const CPUInfo &cpu_info); + + /** Helper function to create the assembly kernel with requantization support + * + * @param[in] src Source tensor info. + * @param[in] dst Destination tensor info. + * @param[in] info Pooling layer meta-data. + */ + template + void create_arm_pooling_requant(const ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &info, const CPUInfo &cpu_info); + + std::unique_ptr _kernel_asm{ nullptr }; +}; +} // namespace kernels +} // namespace cpu +} // namespace arm_compute +#endif /* ARM_COMPUTE_CPU_POOLING_ASSEMBLY_WRAPPER_KERNEL_H */ diff --git a/src/core/cpu/kernels/CpuPoolingKernel.cpp b/src/core/cpu/kernels/CpuPoolingKernel.cpp new file mode 100644 index 0000000000..a29aef4986 --- /dev/null +++ b/src/core/cpu/kernels/CpuPoolingKernel.cpp @@ -0,0 +1,2605 @@ +/* + * Copyright (c) 2017-2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "src/core/cpu/kernels/CpuPoolingKernel.h" + +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/Window.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "src/core/AccessWindowStatic.h" +#include "src/core/CPP/Validate.h" +#include "src/core/NEON/NEAsymm.h" +#include "src/core/NEON/NEFixedPoint.h" +#include "src/core/NEON/NEMath.h" +#include "src/core/helpers/AutoConfiguration.h" +#include "src/core/helpers/WindowHelpers.h" +#include "support/ToolchainSupport.h" + +#include "src/core/NEON/wrapper/wrapper.h" +#include + +namespace arm_compute +{ +namespace cpu +{ +namespace kernels +{ +using namespace misc::shape_calculator; + +namespace +{ +template +inline typename std::enable_if::value, int8_t>::type +quantize(float val, const UniformQuantizationInfo &info) +{ + return quantize_qasymm8_signed(val, info); +} + +template +inline typename std::enable_if::value, uint8_t>::type +quantize(float val, const UniformQuantizationInfo &info) +{ + return quantize_qasymm8(val, info); +} + +inline float calculate_avg_scale(bool exclude_padding, DataLayout data_layout, const Coordinates &id, const int pool_size_x, const int pool_size_y, const int upper_bound_w, const int upper_bound_h, + const int pad_x, const int pad_y, const int stride_x, const int stride_y) +{ + const unsigned int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const unsigned int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + + int start_x = id[idx_width] * stride_x - pad_x; + int start_y = id[idx_height] * stride_y - pad_y; + + const int end_x = std::min(start_x + pool_size_x, upper_bound_w); + const int end_y = std::min(start_y + pool_size_y, upper_bound_h); + if(exclude_padding) + { + start_x = std::max(0, start_x); + start_y = std::max(0, start_y); + } + return 1.f / ((end_y - start_y) * (end_x - start_x)); +} + +template +inline void scale_vector_q16x8(bool exclude_padding, TVec &v, const Coordinates &id, int id_offset, int step, + const int pool_size, const int upper_bound_w, const int upper_bound_h, + const int pad_x, const int pad_y, const int stride_x, const int stride_y) +{ + int start_x = (id.x() + id_offset) * stride_x - pad_x; + int start_y = id.y() * stride_y - pad_y; + const int end_y = std::min(start_y + pool_size, upper_bound_h); + if(exclude_padding) + { + start_y = std::max(0, start_y); + } + + std::array elems = + { + { + wrapper::vgetlane(v, 0), + wrapper::vgetlane(v, 1), + wrapper::vgetlane(v, 2), + wrapper::vgetlane(v, 3), + wrapper::vgetlane(v, 4), + wrapper::vgetlane(v, 5), + wrapper::vgetlane(v, 6), + wrapper::vgetlane(v, 7), + } + }; + + for(auto &el : elems) + { + int c_start_x = start_x; + const int end_x = std::min(c_start_x + pool_size, upper_bound_w); + if(exclude_padding) + { + c_start_x = std::max(0, c_start_x); + } + float scale = 1.f / ((end_y - start_y) * (end_x - c_start_x)); + el *= scale; + start_x += step * stride_x; + } + + v = wrapper::vsetlane(elems[0], v, 0); + v = wrapper::vsetlane(elems[1], v, 1); + v = wrapper::vsetlane(elems[2], v, 2); + v = wrapper::vsetlane(elems[3], v, 3); + v = wrapper::vsetlane(elems[4], v, 4); + v = wrapper::vsetlane(elems[5], v, 5); + v = wrapper::vsetlane(elems[6], v, 6); + v = wrapper::vsetlane(elems[7], v, 7); +} + +Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &pool_info, + unsigned int &pooled_w, unsigned int pooled_h, const ITensorInfo *indices, Size2D pool_size) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); + + int pool_stride_x = 0; + int pool_stride_y = 0; + PoolingType pool_type = pool_info.pool_type; + const PadStrideInfo pad_stride_info = pool_info.pad_stride_info; + std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); + + ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src); + if(indices) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F32, DataType::F16); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(indices, 1, DataType::U32); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(pool_type != PoolingType::MAX, "Pooling indices only supported for MAX pooling method"); + } + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON(pool_type == PoolingType::L2 && is_data_type_quantized(src->data_type())); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_quantized(src->data_type()) && !pool_info.exclude_padding && (pool_info.pool_type == PoolingType::AVG) && pool_info.pad_stride_info.has_padding() + && (src->data_layout() == DataLayout::NHWC), + "exclude_padding equal false is not supported for AVG Pooling with padding on quantized types"); + + if(dst->total_size() != 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, dst); + ARM_COMPUTE_RETURN_ERROR_ON((dst->dimension(get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::WIDTH)) != pooled_w) + || (dst->dimension(get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::HEIGHT)) != pooled_h)); + + if(indices) + { + ARM_COMPUTE_RETURN_ERROR_ON_MSG((pool_size != Size2D(2, 2)), "Pooling indices only supported for pool size 2x2"); + ARM_COMPUTE_RETURN_ERROR_ON((indices->dimension(get_data_layout_dimension_index(indices->data_layout(), DataLayoutDimension::WIDTH)) != pooled_w) + || (indices->dimension(get_data_layout_dimension_index(indices->data_layout(), DataLayoutDimension::HEIGHT)) != pooled_h)); + } + } + + return Status{}; +} + +Status validate_arguments_pool_info(const unsigned int pool_size_x, const unsigned int pool_size_y) +{ + ARM_COMPUTE_RETURN_ERROR_ON(pool_size_x == 0); + ARM_COMPUTE_RETURN_ERROR_ON(pool_size_y == 0); + + return Status{}; +} + +std::pair validate_and_configure_window(ITensorInfo *src, ITensorInfo *dst, ITensorInfo *indices, const PoolingLayerInfo &pool_info, + unsigned int &num_elems_processed_per_iteration, + BorderSize &border_size, + unsigned int pooled_w, unsigned int pooled_h, int pool_size_x, int pool_size_y) +{ + // dst auto inizialitation if not yet initialized + auto_init_if_empty(*dst, src->clone()->set_tensor_shape(compute_pool_shape(*src, pool_info))); + if(indices) + { + // Indices auto inizialitation if not yet initialized + auto_init_if_empty(*indices, (src->clone()->set_tensor_shape(compute_pool_shape(*src, + pool_info))) + .set_data_type(DataType::U32) /* we store the offset to the element */); + } + const auto data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : pool_info.data_layout; + unsigned int num_elems_read_per_iteration = 0; + unsigned int num_elems_horizontal_window = 0; + int pool_stride_x = 0; + int pool_stride_y = 0; + const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + const int src_width = src->dimension(idx_width); + const int src_height = src->dimension(idx_height); + const PadStrideInfo pad_stride_info = pool_info.pad_stride_info; + std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); + const int pool_pad_right = pad_stride_info.pad_right(); + const int pool_pad_top = pad_stride_info.pad_top(); + const int pool_pad_left = pad_stride_info.pad_left(); + const int pool_pad_bottom = pad_stride_info.pad_bottom(); + const bool is_square = pool_size_x == pool_size_y; + + // Check dst dimensions + std::tie(pooled_w, pooled_h) = scaled_dimensions(src->dimension(idx_width), + src->dimension(idx_height), + pool_size_x, + pool_size_y, + pad_stride_info); + + //If it's not squared and optimized will be executed the MxN + num_elems_read_per_iteration = 1; + num_elems_processed_per_iteration = 1; + num_elems_horizontal_window = 1; + + if(is_square) + { + switch(src->data_type()) + { + case DataType::QASYMM8: + case DataType::QASYMM8_SIGNED: + switch(pool_size_x) + { + case 2: + num_elems_read_per_iteration = 16; + num_elems_processed_per_iteration = (pool_stride_x == 2) ? 8 : 15; + num_elems_horizontal_window = (pool_stride_x == 2) ? 8 : 16; + break; + case 3: + num_elems_read_per_iteration = 16; + num_elems_processed_per_iteration = (pool_stride_x == 2) ? 7 : 14; + num_elems_horizontal_window = (pool_stride_x == 2) ? 8 : 16; + break; + default: + break; + } + break; +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + case DataType::F16: + switch(pool_size_x) + { + case 2: + case 3: + num_elems_read_per_iteration = 4; + num_elems_processed_per_iteration = 1; + num_elems_horizontal_window = 1; + break; + default: + break; + } + break; +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + case DataType::F32: + switch(pool_size_x) + { + case 2: + num_elems_read_per_iteration = 2; + break; + case 3: + num_elems_read_per_iteration = 4; // We use vload4 for pooling3 + break; + case 7: + num_elems_read_per_iteration = 8; // We use vload8 for pooling7 + break; + default: + break; + } + num_elems_processed_per_iteration = 1; + num_elems_horizontal_window = 1; + break; + default: + ARM_COMPUTE_ERROR("Element size not supported"); + break; + } + } + + bool window_changed = false; + Window win{}; + if(data_layout == DataLayout::NCHW) + { + // Number of iterations in X dimension + const int num_iterations_x = (pooled_w + num_elems_processed_per_iteration - 1) / num_elems_processed_per_iteration; + // Upper limit for the number of right/bottom border elements that are accessed + const int upper_bound_w = ((num_iterations_x - 1) * num_elems_processed_per_iteration * pool_stride_x - pool_pad_left + num_elems_read_per_iteration) - src_width; + const int upper_bound_h = ((pooled_h - 1) * pool_stride_y - pool_pad_top + pool_size_y) - src_height; + border_size = BorderSize(pool_pad_top, pool_pad_right, pool_pad_bottom, pool_pad_left); + border_size.right = std::max(upper_bound_w, pool_pad_right); + border_size.bottom = std::max(upper_bound_h, pool_pad_bottom); + TensorShape dst_shape{ src->tensor_shape() }; + dst_shape.set(0, pooled_w); + dst_shape.set(1, pooled_h); + TensorInfo dst_info(src->clone()->set_tensor_shape(dst_shape)); + win = calculate_max_window(dst_info, Steps(num_elems_processed_per_iteration)); + AccessWindowStatic src_access(src, -pool_pad_left, -pool_pad_top, src_width + border_size.right, src_height + border_size.bottom); + AccessWindowHorizontal dst_access(dst, 0, num_elems_horizontal_window); + if(indices) + { + AccessWindowHorizontal indices_access(indices, 0, num_elems_horizontal_window); + window_changed = update_window_and_padding(win, src_access, dst_access, indices_access); + } + else + { + window_changed = update_window_and_padding(win, src_access, dst_access); + } + dst_access.set_valid_region(win, ValidRegion(Coordinates(), dst->tensor_shape())); + } + + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; + return std::make_pair(err, win); +} + +template +inline T vcvtq_q32_f32(float32x4_t values); + +template <> +inline uint32x4_t vcvtq_q32_f32(float32x4_t values) +{ + return vcvtq_u32_f32(values); +} + +template <> +inline int32x4_t vcvtq_q32_f32(float32x4_t values) +{ + return vcvtq_s32_f32(values); +} + +template +inline float32x4_t vcvtq_f32_q32(T values); + +template <> +inline float32x4_t vcvtq_f32_q32(uint32x4_t values) +{ + return vcvtq_f32_u32(values); +} + +template <> +inline float32x4_t vcvtq_f32_q32(int32x4_t values) +{ + return vcvtq_f32_s32(values); +} + +template +inline Tout vrequantize_pooling_with_scale(const float32x4x4_t &acc, const float quant_rescale, const float scale_pooling, const int32_t new_offset); + +template <> +inline uint8x16_t vrequantize_pooling_with_scale(const float32x4x4_t &acc, const float quant_rescale, const float scale_pooling, const int32_t new_offset) +{ + const float new_scale = quant_rescale / scale_pooling; + return vquantize(acc, UniformQuantizationInfo(new_scale, new_offset)); +} + +template <> +inline int8x16_t vrequantize_pooling_with_scale(const float32x4x4_t &acc, const float quant_rescale, const float scale_pooling, const int32_t new_offset) +{ + const float new_scale = quant_rescale / scale_pooling; + return vquantize_signed(acc, UniformQuantizationInfo(new_scale, new_offset)); +} + +template +inline Tout vrequantize_pooling(Tin vec1, Tin vec2, const UniformQuantizationInfo &requant_qinfo); + +template <> +inline uint8x16_t vrequantize_pooling(uint8x8_t vec1, uint8x8_t vec2, const UniformQuantizationInfo &requant_qinfo) +{ + const float32x4x4_t acc = + { + { + vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8((vec1))))), + vcvtq_f32_u32(vmovl_u16(vget_high_u16(vmovl_u8((vec1))))), + vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8((vec2))))), + vcvtq_f32_u32(vmovl_u16(vget_high_u16(vmovl_u8((vec2))))), + } + }; + return vquantize(acc, requant_qinfo); +} + +template <> +inline int8x16_t vrequantize_pooling(int8x8_t vec1, int8x8_t vec2, const UniformQuantizationInfo &requant_qinfo) +{ + const float32x4x4_t acc = + { + { + vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8((vec1))))), + vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8((vec1))))), + vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8((vec2))))), + vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8((vec2))))), + } + }; + return vquantize_signed(acc, requant_qinfo); +} + +template +inline T vrequantize_pooling(T &vec, const UniformQuantizationInfo &requant_qinfo); + +template <> +inline uint8x8_t vrequantize_pooling(uint8x8_t &vec, const UniformQuantizationInfo &requant_qinfo) +{ + const float32x4x2_t acc = + { + { + vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8((vec))))), + vcvtq_f32_u32(vmovl_u16(vget_high_u16(vmovl_u8((vec))))), + } + }; + return vquantize(acc, requant_qinfo); +} + +template <> +inline int8x8_t vrequantize_pooling(int8x8_t &vec, const UniformQuantizationInfo &requant_qinfo) +{ + const float32x4x2_t acc = + { + { + vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8((vec))))), + vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8((vec))))), + } + }; + return vquantize_signed(acc, requant_qinfo); +} + +} // namespace + +BorderSize CpuPoolingKernel::border_size() const +{ + return _border_size; +} + +void CpuPoolingKernel::configure(ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &pool_info, ITensorInfo *indices) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); + const PadStrideInfo pad_stride_info = pool_info.pad_stride_info; + const bool is_global_pooling = pool_info.is_global_pooling; + const int pool_stride_x = pad_stride_info.stride().first; + + // Get data layout + const auto data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : pool_info.data_layout; + const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + + // Update pool size in case of global pooling + const Size2D pool_size( + is_global_pooling ? src->dimension(idx_width) : pool_info.pool_size.width, + is_global_pooling ? src->dimension(idx_height) : pool_info.pool_size.height); + + // Validate pool info before calling scaled_dimensions + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_pool_info(pool_size.x(), pool_size.y())); + + // Check dst dimensions + unsigned int pooled_w; + unsigned int pooled_h; + std::tie(pooled_w, pooled_h) = scaled_dimensions(src->dimension(idx_width), + src->dimension(idx_height), + pool_size.x(), + pool_size.y(), + pad_stride_info); + + // Perform validation step + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, pool_info, pooled_w, pooled_h, indices, pool_size)); + + // Set instance variables + _pool_info = pool_info; + _data_layout = src->data_layout(); + _is_square = (pool_size.x() == pool_size.y()); + + // Get data type + const DataType data_type = src->data_type(); + const bool is_nchw = _data_layout == DataLayout::NCHW; + + if(data_type == DataType::QASYMM8) + { + if(!is_nchw) + { + _func = &CpuPoolingKernel::poolingMxN_q8_nhwc; + } + else + { + if(pool_size.x() == 2 && pool_stride_x < 3 && _is_square) + { + _func = &CpuPoolingKernel::pooling2_q8_nchw; + } + else if(pool_size.x() == 3 && pool_stride_x < 3 && _is_square) + { + _func = &CpuPoolingKernel::pooling3_q8_nchw; + } + else + { + _func = &CpuPoolingKernel::poolingMxN_q8_nchw; + } + } + } + else if(data_type == DataType::QASYMM8_SIGNED) + { + if(!is_nchw) + { + _func = &CpuPoolingKernel::poolingMxN_q8_nhwc; + } + else + { + if(pool_size.x() == 2 && pool_stride_x < 3 && _is_square) + { + _func = &CpuPoolingKernel::pooling2_q8_nchw; + } + else if(pool_size.x() == 3 && pool_stride_x < 3 && _is_square) + { + _func = &CpuPoolingKernel::pooling3_q8_nchw; + } + else + { + _func = &CpuPoolingKernel::poolingMxN_q8_nchw; + } + } + } + else if(data_type == DataType::F16) + { + if(!is_nchw) + { + _func = &CpuPoolingKernel::poolingMxN_f16_nhwc; + } + else + { + if(_is_square) + { + switch(pool_size.x()) + { + case 2: + { + _func = &CpuPoolingKernel::pooling2_f16_nchw; + } + break; + case 3: + { + _func = &CpuPoolingKernel::pooling3_f16_nchw; + } + break; + default: + { + _func = &CpuPoolingKernel::poolingMxN_f16_nchw; + break; + } + } + } + else + { + _func = &CpuPoolingKernel::poolingMxN_f16_nchw; + } + } + } + else if(data_type == DataType::F32) + { + if(!is_nchw) + { + _func = &CpuPoolingKernel::poolingMxN_f32_nhwc; + } + else + { + if(_is_square) + { + switch(pool_size.x()) + { + case 2: + { + _func = &CpuPoolingKernel::pooling2_f32_nchw; + break; + } + case 3: + { + _func = &CpuPoolingKernel::pooling3_f32_nchw; + break; + } + case 7: + { + _func = &CpuPoolingKernel::pooling7_f32_nchw; + break; + } + default: + { + _func = &CpuPoolingKernel::poolingMxN_f32_nchw; + break; + } + } + } + else + { + _func = &CpuPoolingKernel::poolingMxN_f32_nchw; + } + } + } + + if(!is_nchw) + { + // Configure kernel window + Window win = calculate_max_window(*dst, Steps()); + Coordinates coord; + coord.set_num_dimensions(dst->num_dimensions()); + dst->set_valid_region(ValidRegion(coord, dst->tensor_shape())); + ICpuKernel::configure(win); + } + else + { + // Configure kernel window + auto win_config = validate_and_configure_window(src, dst, indices, pool_info, _num_elems_processed_per_iteration, + _border_size, pooled_w, pooled_h, pool_size.x(), pool_size.y()); + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + ICpuKernel::configure(win_config.second); + } +} + +template +inline uint32_t offset_no_padding(uint32_t padded_offset, const Coordinates &id, const ITensorInfo &info, int pool_stride_x, int pool_stride_y) +{ + const int pad_left = info.padding().left; + const int pad_right = info.padding().right; + const int pad_top = info.padding().top; + const int pad_bottom = info.padding().bottom; + const int in_stride_y = static_cast(info.strides_in_bytes().y()); + const int in_stride_w = static_cast(info.strides_in_bytes()[3]); + const int pad_horiz = pad_left + pad_right; + const int pad_vert = pad_top + pad_bottom; + + if(info.data_layout() == DataLayout::NCHW) + { + const uint32_t offset_base = padded_offset + - sizeof(T) * pad_horiz * id.y() * pool_stride_y /* subtract padding elems per row */ + - pad_top * sizeof(T) /* top padding */ + - sizeof(T) * pad_horiz * info.tensor_shape()[1] * id.z() - pad_vert * in_stride_y * id.z() /* for each Z plane there are height*pad_right padding elems */ + - in_stride_w * id[3]; + + return offset_base; + } + else + { + const uint32_t offset_base = padded_offset + - sizeof(T) * pad_horiz * id.y() * pool_stride_x // subtract padding elems per row + - pad_top * sizeof(T) // top padding + - sizeof(T) * pad_horiz * info.tensor_shape()[1] * id.z() * pool_stride_y // for each Z plane there are width*pad_right padding elems + - in_stride_w * id[3]; + + return offset_base; + } +} + +template +void CpuPoolingKernel::pooling2_q8_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) +{ + Iterator src(_src, window_src); + Iterator dst(_dst, window); + + /** NEON vector types */ + using q8x8_t = typename wrapper::traits::neon_vector::type; + using q8x16_t = typename wrapper::traits::neon_vector::type; + using q8x8x2_t = typename std::conditional::value, uint8x8x2_t, int8x8x2_t>::type; + using q16_t = typename wrapper::traits::promote_t; + using q16x4_t = typename wrapper::traits::neon_vector::type; + using q16x8_t = typename wrapper::traits::neon_vector::type; + using q16x8x2_t = typename wrapper::traits::neon_vector::type; + + constexpr int pool_size = 2; + int pool_stride_x = 0; + int pool_stride_y = 0; + const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); + std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); + const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); + + const T *const src_top_ptr = reinterpret_cast(_src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top)))); + const T *const src_bottom_ptr = reinterpret_cast(_src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1))); + + const int scale_step_x = (pool_stride_x == 1) ? 2 : 1; + + const UniformQuantizationInfo src_qinfo = _src->info()->quantization_info().uniform(); + const UniformQuantizationInfo dst_qinfo = _dst->info()->quantization_info().uniform(); + const bool have_different_qinfo = src_qinfo != dst_qinfo; + + const float requant_scale = dst_qinfo.scale / src_qinfo.scale; + const int32_t requant_offset = dst_qinfo.offset - static_cast(static_cast(src_qinfo.offset) / requant_scale); + const UniformQuantizationInfo requant_qinfo = UniformQuantizationInfo(requant_scale, requant_offset); + + execute_window_loop(window, [&](const Coordinates & id) + { + const auto top_data = wrapper::vloadq(src_top_ptr + src.offset()); + const auto bottom_data = wrapper::vloadq(src_bottom_ptr + src.offset()); + q8x8_t lower_res = {}; + q8x8_t upper_res = {}; + + if(pooling_type != PoolingType::MAX) + { + const q16x8x2_t top_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(top_data)), wrapper::vmovl(wrapper::vgethigh(top_data)) } }; + const q16x8x2_t bottom_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(bottom_data)), wrapper::vmovl(wrapper::vgethigh(bottom_data)) } }; + + // Add rows + const q16x8x2_t vrsum = + { + { + wrapper::vadd(top_data_q16.val[0], bottom_data_q16.val[0]), + wrapper::vadd(top_data_q16.val[1], bottom_data_q16.val[1]), + } + }; + + // Pair-wise add row data + const q16x4_t vpsum_1 = wrapper::vpadd(wrapper::vgetlow(vrsum.val[0]), wrapper::vgethigh(vrsum.val[0])); + const q16x4_t vpsum_2 = wrapper::vpadd(wrapper::vgetlow(vrsum.val[1]), wrapper::vgethigh(vrsum.val[1])); + + q16x8_t res_lower = wrapper::vcombine(vpsum_1, vpsum_2); + + // Scale lower result + scale_vector_q16x8(exclude_padding, res_lower, id, 0, scale_step_x, + pool_size, upper_bound_w, upper_bound_h, + pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + lower_res = wrapper::vmovn(res_lower); + + // Compute upper result for stride_x == 1 + if(pool_stride_x == 1) + { + // Shifted row sum + const q16x8x2_t vrsum_shifted = + { + { + wrapper::vext_1(vrsum.val[0], vrsum.val[1]), + wrapper::vext_1(vrsum.val[1], vrsum.val[1]) + } + }; + + // Pair-wise add shifted row + q16x8_t res_upper = wrapper::vcombine( + wrapper::vpadd(wrapper::vgetlow(vrsum_shifted.val[0]), wrapper::vgethigh(vrsum_shifted.val[0])), + wrapper::vpadd(wrapper::vgetlow(vrsum_shifted.val[1]), wrapper::vgethigh(vrsum_shifted.val[1]))); + + // Scale upper result + scale_vector_q16x8(exclude_padding, res_upper, id, 1, 2, + pool_size, upper_bound_w, upper_bound_h, + pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + upper_res = wrapper::vmovn(res_upper); + } + } + else + { + const q8x16_t max_data = wrapper::vmax(top_data, bottom_data); + lower_res = wrapper::vpmax(wrapper::vgetlow(max_data), wrapper::vgethigh(max_data)); + if(pool_stride_x == 1) + { + const q8x16_t max_data_shifted = wrapper::vext_1(max_data, max_data); + upper_res = wrapper::vpmax(wrapper::vgetlow(max_data_shifted), wrapper::vgethigh(max_data_shifted)); + } + } + + if(have_different_qinfo) + { + const auto requantized_dst = vrequantize_pooling(lower_res, upper_res, requant_qinfo); + lower_res = wrapper::vgetlow(requantized_dst); + upper_res = wrapper::vgethigh(requantized_dst); + } + + // Store result + if(pool_stride_x == 1) + { + const q8x8x2_t res = { { lower_res, upper_res } }; + wrapper::vstore(reinterpret_cast(dst.ptr()), res); + } + else + { + wrapper::vstore(reinterpret_cast(dst.ptr()), lower_res); + } + }, + src, dst); +} + +void CpuPoolingKernel::pooling3_f16_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) +{ + ARM_COMPUTE_UNUSED(pooling_type); + ARM_COMPUTE_UNUSED(exclude_padding); +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + Iterator src(_src, window_src); + Iterator dst(_dst, window); + + constexpr const int pool_size = 3; + const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); + const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); + + const unsigned char *const src_top_ptr = _src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top))); + const unsigned char *const src_middle_ptr = _src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1)); + const unsigned char *const src_bottom_ptr = _src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 2)); + + execute_window_loop(window, [&](const Coordinates & id) + { + float16x4_t top_data = vld1_f16(reinterpret_cast(src_top_ptr + src.offset())); + float16x4_t middle_data = vld1_f16(reinterpret_cast(src_middle_ptr + src.offset())); + float16x4_t bottom_data = vld1_f16(reinterpret_cast(src_bottom_ptr + src.offset())); + float16x4_t res = {}; + + // Get power of 2 in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + top_data = vmul_f16(top_data, top_data); + middle_data = vmul_f16(middle_data, middle_data); + bottom_data = vmul_f16(bottom_data, bottom_data); + } + + if(pooling_type != PoolingType::MAX) + { + // Calculate scale + const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + const float16x4_t scale_v = vdup_n_f16(scale); + // Perform pooling + const float16x4_t sum_data = vadd_f16(vadd_f16(top_data, bottom_data), middle_data); + res = vpadd_f16(vset_lane_f16(0.f, sum_data, 3), sum_data); + res = vmul_f16(vpadd_f16(res, res), scale_v); + } + else + { + const float16x4_t max_data = vmax_f16(vmax_f16(top_data, bottom_data), middle_data); + res = vpmax_f16(vset_lane_f16(-std::numeric_limits::max(), max_data, 3), max_data); + res = vpmax_f16(res, res); + } + + // Calculate square-root in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + res = vinv_f16(vinvsqrt_f16(res)); + } + + *(reinterpret_cast(dst.ptr())) = vget_lane_f16(res, 0); + }, + src, dst); +#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + ARM_COMPUTE_UNUSED(window_src); + ARM_COMPUTE_UNUSED(window); + ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ +} + +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +template +inline typename std::enable_if::value, float32x2_t>::type +f16_to_f32(float16x4_t src) +{ + float32x2_t dst = { static_cast(vget_lane_f16(src, 0)), static_cast(vget_lane_f16(src, 1)) }; + return dst; +} +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + +template +inline typename std::enable_if::value, float32x2_t>::type +f16_to_f32(float32x2_t src) +{ + return src; +} + +template +void CpuPoolingKernel::pooling2_nchw_maxpool_indices(const Window &window_src, const Window &window) +{ + Iterator src(_src, window_src); + Iterator dst(_dst, window); + Iterator indices(_indices, window); + const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); + const uint8_t *const src_top_ptr = _src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top))); + const uint8_t *const src_bottom_ptr = _src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1)); + const int pad_left = _src->info()->padding().left; + const int pad_right = _src->info()->padding().right; + const int in_stride_y = static_cast(_src->info()->strides_in_bytes().y()); + + execute_window_loop(window, [&](const Coordinates & id) + { + auto top_data = wrapper::vload(reinterpret_cast(src_top_ptr + src.offset())); + auto bottom_data = wrapper::vload(reinterpret_cast(src_bottom_ptr + src.offset())); + float32x2_t top_data_f32 = f16_to_f32(top_data); + float32x2_t bottom_data_f32 = f16_to_f32(bottom_data); + + // Calculate max data, compare top first, then bottom, to make sue the first max is recorded. + const float32x2_t max_data_top = vpmax_f32(top_data_f32, top_data_f32); + const float32x2_t max_data_bottom = vpmax_f32(bottom_data_f32, bottom_data_f32); + const float32x2_t max_data = vmax_f32(max_data_top, max_data_bottom); + *(reinterpret_cast(dst.ptr())) = static_cast(vget_lane_f32(max_data, 0)); + + // Calculate max data indice, which will be used in max unpool. + const uint32_t offset_base = offset_no_padding(src.offset(), id, *_src->info(), pool_stride_x, pool_stride_y); + const uint32_t offset_top = (uint32_t)(offset_base / sizeof(T)); + const uint32_t offset_bottom = offset_top + in_stride_y / sizeof(T) - pad_right - pad_left; + const uint32x2_t voffset_top = { offset_top, offset_top + 1u }; + const uint32x2_t voffset_bottom = { offset_bottom, offset_bottom + 1u }; + const uint32x2_t tmp_indices_top = vbsl_u32(vcge_f32(top_data_f32, vrev64_f32(top_data_f32)), voffset_top, vrev64_u32(voffset_top)); + const uint32x2_t tmp_indices_bottom = vbsl_u32(vcge_f32(bottom_data_f32, vrev64_f32(bottom_data_f32)), voffset_bottom, vrev64_u32(voffset_bottom)); + *(reinterpret_cast(indices.ptr())) = vget_lane_u32(vbsl_u32(vcge_f32(max_data_top, max_data_bottom), tmp_indices_top, tmp_indices_bottom), 0); + }, + src, dst, indices); +} + +void CpuPoolingKernel::pooling2_f16_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) +{ + ARM_COMPUTE_UNUSED(pooling_type); + ARM_COMPUTE_UNUSED(exclude_padding); +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + if(pooling_type == PoolingType::MAX && _indices) + { + pooling2_nchw_maxpool_indices(window_src, window); + } + else + { + Iterator src(_src, window_src); + Iterator dst(_dst, window); + constexpr int pool_size = 2; + const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); + int pool_stride_x, pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); + const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); + + const unsigned char *const src_top_ptr = _src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top))); + const unsigned char *const src_bottom_ptr = _src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1)); + + execute_window_loop(window, [&](const Coordinates & id) + { + float16x4_t top_data = vld1_f16(reinterpret_cast(src_top_ptr + src.offset())); + float16x4_t bottom_data = vld1_f16(reinterpret_cast(src_bottom_ptr + src.offset())); + float16x4_t res = {}; + + // Get power of 2 in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + top_data = vmul_f16(top_data, top_data); + bottom_data = vmul_f16(bottom_data, bottom_data); + } + + if(pooling_type != PoolingType::MAX) + { + const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + const float16x4_t scale_v = vdup_n_f16(scale); + + const float16x4_t sum_data = vadd_f16(top_data, bottom_data); + res = vmul_f16(vpadd_f16(sum_data, sum_data), scale_v); + } + else + { + const float16x4_t max_data = vmax_f16(top_data, bottom_data); + res = vpmax_f16(max_data, max_data); + } + + // Calculate square-root in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + res = vinv_f16(vinvsqrt_f16(res)); + } + + // Store result + *(reinterpret_cast(dst.ptr())) = vget_lane_f16(res, 0); + }, + src, dst); + } +#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + ARM_COMPUTE_UNUSED(window_src); + ARM_COMPUTE_UNUSED(window); + ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ +} + +template +void CpuPoolingKernel::pooling3_q8_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) +{ + Iterator src(_src, window_src); + Iterator dst(_dst, window); + + /** NEON vector types */ + using q8x8_t = typename wrapper::traits::neon_vector::type; + using q8x16_t = typename wrapper::traits::neon_vector::type; + using q8x8x2_t = typename std::conditional::value, uint8x8x2_t, int8x8x2_t>::type; + using q16_t = typename wrapper::traits::promote_t; + using q16x8_t = typename wrapper::traits::neon_vector::type; + using q16x8x2_t = typename wrapper::traits::neon_vector::type; + + constexpr int pool_size = 3; + const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); + const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); + + const UniformQuantizationInfo &src_qinfo = _src->info()->quantization_info().uniform(); + const UniformQuantizationInfo &dst_qinfo = _dst->info()->quantization_info().uniform(); + + const float requant_scale = dst_qinfo.scale / src_qinfo.scale; + const int32_t requant_offset = dst_qinfo.offset - static_cast(static_cast(src_qinfo.offset) / requant_scale); + const UniformQuantizationInfo requant_qinfo = UniformQuantizationInfo(requant_scale, requant_offset); + + const T *const src_top_ptr = reinterpret_cast(_src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top)))); + const T *const src_middle_ptr = reinterpret_cast(_src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1))); + const T *const src_bottom_ptr = reinterpret_cast(_src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 2))); + + execute_window_loop(window, [&](const Coordinates & id) + { + const auto top_data = wrapper::vloadq(src_top_ptr + src.offset()); + const auto middle_data = wrapper::vloadq(src_middle_ptr + src.offset()); + const auto bottom_data = wrapper::vloadq(src_bottom_ptr + src.offset()); + q8x8_t fres = {}; + q8x16_t fqres = {}; + + if(pooling_type == PoolingType::AVG) + { + // Convert data to u16 + const q16x8x2_t top_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(top_data)), wrapper::vmovl(wrapper::vgethigh(top_data)) } }; + const q16x8x2_t middle_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(middle_data)), wrapper::vmovl(wrapper::vgethigh(middle_data)) } }; + const q16x8x2_t bottom_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(bottom_data)), wrapper::vmovl(wrapper::vgethigh(bottom_data)) } }; + + // Calculate row sums + const q16x8x2_t vrsum = + { + { + wrapper::vadd(wrapper::vadd(top_data_q16.val[0], bottom_data_q16.val[0]), middle_data_q16.val[0]), + wrapper::vadd(wrapper::vadd(top_data_q16.val[1], bottom_data_q16.val[1]), middle_data_q16.val[1]), + } + }; + const q16x8x2_t vrsum_shifted_1 = + { + { + wrapper::vext_1(vrsum.val[0], vrsum.val[1]), + wrapper::vext_1(vrsum.val[1], vrsum.val[1]) + } + }; + const q16x8x2_t vrsum_shifted_2 = + { + { + wrapper::vext_2(vrsum.val[0], vrsum.val[1]), + wrapper::vext_2(vrsum.val[1], vrsum.val[1]) + } + }; + // Calculate final sum + q16x8x2_t final_sum = + { + { + wrapper::vadd(wrapper::vadd(vrsum.val[0], vrsum_shifted_1.val[0]), vrsum_shifted_2.val[0]), + wrapper::vadd(wrapper::vadd(vrsum.val[1], vrsum_shifted_1.val[1]), vrsum_shifted_2.val[1]), + } + }; + if(pool_stride_x == 2) + { + q16x8_t res = + { + wrapper::vgetlane(final_sum.val[0], 0), + wrapper::vgetlane(final_sum.val[0], 2), + wrapper::vgetlane(final_sum.val[0], 4), + wrapper::vgetlane(final_sum.val[0], 6), + wrapper::vgetlane(final_sum.val[1], 0), + wrapper::vgetlane(final_sum.val[1], 2), + wrapper::vgetlane(final_sum.val[1], 4), + wrapper::vgetlane(final_sum.val[1], 6), + }; + + scale_vector_q16x8(exclude_padding, res, id, 0, 1, + pool_size, upper_bound_w, upper_bound_h, + pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + fres = wrapper::vmovn(res); + } + else + { + // Scale lower result + scale_vector_q16x8(exclude_padding, final_sum.val[0], id, 0, 1, + pool_size, upper_bound_w, upper_bound_h, + pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + // Scale lower result + scale_vector_q16x8(exclude_padding, final_sum.val[1], id, 8, 1, + pool_size, upper_bound_w, upper_bound_h, + pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + fqres = wrapper::vcombine(wrapper::vmovn(final_sum.val[0]), wrapper::vmovn(final_sum.val[1])); + } + } + else + { + const q8x16_t max_data = wrapper::vmax(wrapper::vmax(top_data, bottom_data), middle_data); + const q8x16_t max_data_shift1 = wrapper::vext_1(max_data, max_data); + const q8x16_t max_data_shift2 = wrapper::vext_2(max_data, max_data); + const q8x16_t final_max = wrapper::vmax(wrapper::vmax(max_data, max_data_shift1), max_data_shift2); + + if(pool_stride_x == 2) + { + const q8x8x2_t table = { { wrapper::vgetlow(final_max), wrapper::vgethigh(final_max) } }; + static const q8x8_t lookup_val = { 0, 2, 4, 6, 8, 10, 12, 14 }; + fres = wrapper::vtbl(table, lookup_val); + } + else + { + fqres = final_max; + } + } + + // Store result + if(pool_stride_x == 1) + { + if(src_qinfo != dst_qinfo) + { + fqres = vrequantize_pooling(wrapper::vgetlow(fqres), wrapper::vgethigh(fqres), requant_qinfo); + } + wrapper::vstore(reinterpret_cast(dst.ptr()), fqres); + } + else + { + if(src_qinfo != dst_qinfo) + { + fres = vrequantize_pooling(fres, requant_qinfo); + } + wrapper::vstore(reinterpret_cast(dst.ptr()), fres); + } + }, + src, dst); +} + +void CpuPoolingKernel::poolingMxN_f16_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) +{ + ARM_COMPUTE_UNUSED(pooling_type); + ARM_COMPUTE_UNUSED(exclude_padding); +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + Iterator src(_src, window_src); + Iterator dst(_dst, window); + + const int pool_size_x = _pool_info.is_global_pooling ? _src->info()->tensor_shape().x() : _pool_info.pool_size.width; + const int pool_size_y = _pool_info.is_global_pooling ? _src->info()->tensor_shape().y() : _pool_info.pool_size.height; + const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); + const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); + + execute_window_loop(window, [&](const Coordinates & id) + { + float16_t res = 0.0f; + float16x8_t vres = vdupq_n_f16(0.0f); + + if(pooling_type != PoolingType::MAX) + { + // Calculate scale + const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + + // Perform pooling + + for(int y = 0; y < pool_size_y; ++y) + { + int x = 0; + for(; x <= (pool_size_x - 8); x += 8) + { + const float16x8_t data = vld1q_f16(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().y()))); + + // Get power of 2 in case of l2 pooling and accumulate + if(pooling_type == PoolingType::L2) + { + vres = vaddq_f16(vres, vmulq_f16(data, data)); + } + else + { + vres = vaddq_f16(vres, data); + } + } + + // Leftover for loop + for(; x < pool_size_x; ++x) + { + float16_t data = *(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().x()) + + (y - pool_pad_top) * static_cast(_src->info()->strides_in_bytes().y()))); + + // Get power of 2 in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + data *= data; + } + + res += data; + } + } + + // Reduction + float16x4_t tmp = vpadd_f16(vget_high_f16(vres), vget_low_f16(vres)); + res += vget_lane_f16(tmp, 0); + res += vget_lane_f16(tmp, 1); + res += vget_lane_f16(tmp, 2); + res += vget_lane_f16(tmp, 3); + + // Divide by scale + res *= scale; + } + else + { + float16x8_t vres = vdupq_n_f16(std::numeric_limits::lowest()); + res = std::numeric_limits::lowest(); + + for(int y = 0; y < pool_size_y; ++y) + { + int x = 0; + for(; x <= (pool_size_x - 8); x += 8) + { + const float16x8_t data = vld1q_f16(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().y()))); + vres = vmaxq_f16(vres, data); + } + + // Leftover for loop + for(; x < pool_size_x; ++x) + { + const float16_t data = *(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().x()) + + (y - pool_pad_top) * static_cast(_src->info()->strides_in_bytes().y()))); + res = std::max(res, data); + } + } + + float16x4_t tmp = vpmax_f16(vget_high_f16(vres), vget_low_f16(vres)); + res = std::max(res, vget_lane_f16(tmp, 0)); + res = std::max(res, vget_lane_f16(tmp, 1)); + res = std::max(res, vget_lane_f16(tmp, 2)); + res = std::max(res, vget_lane_f16(tmp, 3)); + } + + // Calculate square-root in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + res = std::sqrt(res); + } + + // Store result + *(reinterpret_cast(dst.ptr())) = res; + }, + src, dst); + +#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + ARM_COMPUTE_UNUSED(window_src); + ARM_COMPUTE_UNUSED(window); + ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ +} + +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +void CpuPoolingKernel::pooling2_f16_nhwc_maxpool_indices(const Window &window_src, const Window &window) +{ + const int window_start_x = window.x().start(); + const int window_end_x = window.x().end(); + const int window_step_x = 8; + + Window window_out = window; + window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator src(_src, window_src); + Iterator dst(_dst, window_out); + Iterator indices(_indices, window_out); + + const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); + + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); + + const int pad_right = _src->info()->padding().right; + const int in_stride_y = static_cast(_src->info()->strides_in_bytes().y()); + const int in_stride_z = static_cast(_src->info()->strides_in_bytes().z()); + + execute_window_loop(window_out, [&](const Coordinates & id) + { + const int idx_width = id.y() * pool_stride_x; + const int idx_height = id.z() * pool_stride_y; + const int pool_limit_y = pool_pad_top - idx_height; + const int pool_limit_x = pool_pad_left - idx_width; + + const int pool_start_y = std::max(0, window_src.z().start() + pool_limit_y); + const int pool_start_x = std::max(0, window_src.y().start() + pool_limit_x); + const int in_x0_offset = (pool_start_x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (pool_start_y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z()); + const int in_x1_offset = (pool_start_x + 1 - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (pool_start_y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z()); + const int in_x2_offset = (pool_start_x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (pool_start_y + 1 - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z()); + const int in_x3_offset = (pool_start_x + 1 - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (pool_start_y + 1 - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z()); + + int x_off = window_start_x; + for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) + { + const auto in_x0_ptr = reinterpret_cast(src.ptr() + in_x0_offset) + x_off; + const auto in_x1_ptr = reinterpret_cast(src.ptr() + in_x1_offset) + x_off; + const auto in_x2_ptr = reinterpret_cast(src.ptr() + in_x2_offset) + x_off; + const auto in_x3_ptr = reinterpret_cast(src.ptr() + in_x3_offset) + x_off; + const auto v_x0 = vld1q_f16(in_x0_ptr); + const auto v_x1 = vld1q_f16(in_x1_ptr); + const auto v_x2 = vld1q_f16(in_x2_ptr); + const auto v_x3 = vld1q_f16(in_x3_ptr); + float16x8_t vres = vmaxq_f16(vmaxq_f16(v_x2, v_x3), vmaxq_f16(v_x0, v_x1)); + // Store result + vst1q_f16(reinterpret_cast(dst.ptr()) + x_off, vres); + + const uint32_t offset_base = offset_no_padding(src.offset(), id, *_src->info(), pool_stride_x, pool_stride_y); + const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float16_t) + x_off; + const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float16_t) - pad_right; + const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float16_t) - pad_right * _src->info()->tensor_shape()[1]; + const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float16_t) - pad_right; + const uint32x4_t voffset_x0_0 = { offset_x0, offset_x0 + 1, offset_x0 + 2, offset_x0 + 3 }; + const uint32x4_t voffset_x0_1 = { offset_x0 + 4, offset_x0 + 5, offset_x0 + 6, offset_x0 + 7 }; + const uint16x8_t voffset_x0 = vcombine_u16(vmovn_u32(voffset_x0_0), vmovn_u32(voffset_x0_1)); + const uint32x4_t voffset_x1_0 = { offset_x1, offset_x1 + 1, offset_x1 + 2, offset_x1 + 3 }; + const uint32x4_t voffset_x1_1 = { offset_x1 + 4, offset_x1 + 5, offset_x1 + 6, offset_x1 + 7 }; + const uint16x8_t voffset_x1 = vcombine_u16(vmovn_u32(voffset_x1_0), vmovn_u32(voffset_x1_1)); + const uint32x4_t voffset_x2_0 = { offset_x2, offset_x2 + 1, offset_x2 + 2, offset_x2 + 3 }; + const uint32x4_t voffset_x2_1 = { offset_x2 + 4, offset_x2 + 5, offset_x2 + 6, offset_x2 + 7 }; + const uint16x8_t voffset_x2 = vcombine_u16(vmovn_u32(voffset_x2_0), vmovn_u32(voffset_x2_1)); + const uint32x4_t voffset_x3_0 = { offset_x3, offset_x3 + 1, offset_x3 + 2, offset_x3 + 3 }; + const uint32x4_t voffset_x3_1 = { offset_x3 + 4, offset_x3 + 5, offset_x3 + 6, offset_x3 + 7 }; + const uint16x8_t voffset_x3 = vcombine_u16(vmovn_u32(voffset_x3_0), vmovn_u32(voffset_x3_1)); + const uint16x8_t tmp_indices0 = vbslq_u16(vcgeq_f16(v_x0, v_x1), voffset_x0, voffset_x1); + const uint16x8_t tmp_indices1 = vbslq_u16(vcgeq_f16(v_x2, v_x3), voffset_x2, voffset_x3); + const uint16x8_t tmp_indices2 = vbslq_u16(vcgeq_f16(vmaxq_f16(v_x0, v_x1), vmaxq_f16(v_x2, v_x3)), tmp_indices0, tmp_indices1); + const uint32x4_t tmp_indeces3_0 = vmovl_u16(vget_low_u16(tmp_indices2)); + const uint32x4_t tmp_indeces3_1 = vmovl_u16(vget_high_u16(tmp_indices2)); + // Store indicies + vst1q_u32(reinterpret_cast(indices.ptr()) + x_off, tmp_indeces3_0); + vst1q_u32(reinterpret_cast(indices.ptr() + 16) + x_off, tmp_indeces3_1); + } + + // Left-overs loop + for(; x_off < window_end_x; ++x_off) + { + const auto x0 = *(reinterpret_cast(src.ptr() + in_x0_offset) + x_off); + const auto x1 = *(reinterpret_cast(src.ptr() + in_x1_offset) + x_off); + const auto x2 = *(reinterpret_cast(src.ptr() + in_x2_offset) + x_off); + const auto x3 = *(reinterpret_cast(src.ptr() + in_x3_offset) + x_off); + float16_t res = std::max(std::max(x2, x3), std::max(x0, x1)); + + // Store result + *(reinterpret_cast(dst.ptr()) + x_off) = res; + + const uint32_t offset_base = offset_no_padding(src.offset(), id, *_src->info(), pool_stride_x, pool_stride_y); + const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float16_t) + x_off; + const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float16_t) - pad_right; + const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float16_t) - pad_right * _src->info()->tensor_shape()[1]; + const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float16_t) - pad_right; + const uint32_t tmp_idx0 = (x0 >= x1) ? offset_x0 : offset_x1; + const uint32_t tmp_idx1 = (x2 >= x3) ? offset_x2 : offset_x3; + const uint32_t tmp_idx2 = (std::max(x0, x1) >= std::max(x2, x3)) ? tmp_idx0 : tmp_idx1; + + // Store indices + *(reinterpret_cast(indices.ptr()) + x_off) = tmp_idx2; + } + }, + src, dst, indices); +} +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + +void CpuPoolingKernel::poolingMxN_f16_nhwc(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) +{ + ARM_COMPUTE_UNUSED(pooling_type); + ARM_COMPUTE_UNUSED(exclude_padding); +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + if(_pool_info.pool_size == Size2D(2, 2) && pooling_type == PoolingType::MAX && _indices) + { + pooling2_f16_nhwc_maxpool_indices(window_src, window); + } + const int window_start_x = window.x().start(); + const int window_end_x = window.x().end(); + const int window_step_x = 8; + + Window window_out = window; + window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator src(_src, window_src); + Iterator dst(_dst, window_out); + + const int pool_size_x = _pool_info.is_global_pooling ? _src->info()->tensor_shape().y() : _pool_info.pool_size.width; + const int pool_size_y = _pool_info.is_global_pooling ? _src->info()->tensor_shape().z() : _pool_info.pool_size.height; + const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); + const int upper_bound_w = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = _src->info()->dimension(2) + (exclude_padding ? 0 : pool_pad_bottom); + + float16x8_t vres; + + execute_window_loop(window_out, [&](const Coordinates & id) + { + const int idx_width = id.y() * pool_stride_x; + const int idx_height = id.z() * pool_stride_y; + const int pool_limit_y = pool_pad_top - idx_height; + const int pool_limit_x = pool_pad_left - idx_width; + + const int pool_start_y = std::max(0, window_src.z().start() + pool_limit_y); + const int pool_end_y = std::min(pool_size_y, window_src.z().end() + pool_limit_y); + const int pool_start_x = std::max(0, window_src.y().start() + pool_limit_x); + const int pool_end_x = std::min(pool_size_x, window_src.y().end() + pool_limit_x); + + int x_off = window_start_x; + for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) + { + if(pooling_type != PoolingType::MAX) + { + // Calculate scale + const float scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, + pool_stride_y); + const float16x8_t scale_v = vdupq_n_f16(scale); + + // Perform pooling + vres = vdupq_n_f16(0.0f); + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const float16x8_t data = vld1q_f16(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z())) + x_off); + + // Get power of 2 in case of l2 pooling and accumulate + if(pooling_type == PoolingType::L2) + { + vres = vaddq_f16(vres, vmulq_f16(data, data)); + } + else + { + vres = vaddq_f16(vres, data); + } + } + } + // Divide by scale + vres = vmulq_f16(vres, scale_v); + } + else + { + vres = vdupq_n_f16(std::numeric_limits::lowest()); + + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const float16x8_t data = vld1q_f16(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z())) + x_off); + vres = vmaxq_f16(vres, data); + } + } + } + + // Calculate square-root in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + float16x8_t sqrt_reciprocal = vrsqrteq_f16(vres); + vres = vmulq_f16(vres, vmulq_f16(vrsqrtsq_f16(vmulq_f16(vres, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal)); + } + + // Store result + vst1q_f16(reinterpret_cast(dst.ptr()) + x_off, vres); + } + + // Left-overs loop + for(; x_off < window_end_x; ++x_off) + { + float16_t res = 0.0f; + + if(pooling_type != PoolingType::MAX) + { + // Calculate scale + const float16_t scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, + pool_stride_y); + + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const float data = *(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z())) + x_off); + + // Get power of 2 in case of l2 pooling and accumulate + if(pooling_type == PoolingType::L2) + { + res += data * data; + } + else + { + res += data; + } + } + } + + // Divide by scale + res *= scale; + } + else + { + res = std::numeric_limits::lowest(); + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const float16_t data = *(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z())) + x_off); + res = std::max(res, data); + } + } + } + + // Calculate square-root in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + res = std::sqrt(res); + } + + // Store result + *(reinterpret_cast(dst.ptr()) + x_off) = res; + } + }, + src, dst); + +#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + ARM_COMPUTE_UNUSED(window_src); + ARM_COMPUTE_UNUSED(window); + ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ +} + +void CpuPoolingKernel::poolingMxN_f32_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) +{ + Iterator src(_src, window_src); + Iterator dst(_dst, window); + + const int pool_size_x = _pool_info.is_global_pooling ? _src->info()->tensor_shape().x() : _pool_info.pool_size.width; + const int pool_size_y = _pool_info.is_global_pooling ? _src->info()->tensor_shape().y() : _pool_info.pool_size.height; + const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); + const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); + + execute_window_loop(window, [&](const Coordinates & id) + { + float res = 0.0f; + + if(pooling_type != PoolingType::MAX) + { + // Calculate scale + const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + + // Perform pooling + float32x4_t vres = vdupq_n_f32(0.0f); + + for(int y = 0; y < pool_size_y; ++y) + { + int x = 0; + for(; x <= (pool_size_x - 4); x += 4) + { + const float32x4_t data = vld1q_f32(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().y()))); + + // Get power of 2 in case of l2 pooling and accumulate + if(pooling_type == PoolingType::L2) + { + vres = vmlaq_f32(vres, data, data); + } + else + { + vres = vaddq_f32(vres, data); + } + } + + // Leftover for loop + for(; x < pool_size_x; ++x) + { + float data = *(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().y()))); + + // Get power of 2 in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + data *= data; + } + + res += data; + } + } + +#if defined(__aarch64__) + // Reduction operation available on 64 bit architectures only + res += vaddvq_f32(vres); +#else // __aarch64__ + // Reduction + float32x2_t tmp = vpadd_f32(vget_high_f32(vres), vget_low_f32(vres)); + tmp = vpadd_f32(tmp, tmp); + + res += vget_lane_f32(tmp, 0); +#endif // __aarch64__ + // Divide by scale + res *= scale; + } + else + { + float32x4_t vres = vdupq_n_f32(std::numeric_limits::lowest()); + res = std::numeric_limits::lowest(); + + for(int y = 0; y < pool_size_y; ++y) + { + int x = 0; + for(; x <= (pool_size_x - 4); x += 4) + { + const float32x4_t data = vld1q_f32(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().y()))); + vres = vmaxq_f32(vres, data); + } + + // Leftover for loop + for(; x < pool_size_x; ++x) + { + const float data = *(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().y()))); + res = std::max(res, data); + } + } +#if defined(__aarch64__) + // Reduction operation available on 64 bit architectures only + res = std::max(vmaxvq_f32(vres), res); +#else // __aarch64__ + float32x2_t tmp = vpmax_f32(vget_high_f32(vres), vget_low_f32(vres)); + tmp = vpmax_f32(tmp, tmp); + + res = std::max(res, vget_lane_f32(tmp, 0)); +#endif // __aarch64__ + } + + // Calculate square-root in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + res = std::sqrt(res); + } + + // Store result + *(reinterpret_cast(dst.ptr())) = res; + }, + src, dst); +} + +void CpuPoolingKernel::pooling2_f32_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, + bool exclude_padding) +{ + if(pooling_type == PoolingType::MAX && _indices) + { + pooling2_nchw_maxpool_indices(window_src, window); + } + else + { + Iterator src(_src, window_src); + Iterator dst(_dst, window); + constexpr int pool_size = 2; + const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); + const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); + + const uint8_t *const src_top_ptr = _src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top))); + const uint8_t *const src_bottom_ptr = _src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1)); + + execute_window_loop(window, [&](const Coordinates & id) + { + const auto in_top_ptr = reinterpret_cast(src_top_ptr + src.offset()); + const auto in_bottom_ptr = reinterpret_cast(src_bottom_ptr + src.offset()); + float32x2_t top_data = vld1_f32(in_top_ptr); + float32x2_t bottom_data = vld1_f32(in_bottom_ptr); + float32x2_t res = {}; + float final_res = 0; + // Get power of 2 in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + top_data = vmul_f32(top_data, top_data); + bottom_data = vmul_f32(bottom_data, bottom_data); + } + + if(pooling_type != PoolingType::MAX) + { + // Calculate scale + float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + const float32x2_t scale_v = vdup_n_f32(scale); + + // Perform pooling + const float32x2_t sum_data = vadd_f32(top_data, bottom_data); + res = vmul_f32(vpadd_f32(sum_data, sum_data), scale_v); + } + else + { + const float32x2_t max_data = vmax_f32(top_data, bottom_data); + res = vpmax_f32(max_data, max_data); + } + final_res = vget_lane_f32(res, 0); + + // Calculate square-root in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + final_res = sqrt(final_res); + } + + // Store result + *(reinterpret_cast(dst.ptr())) = final_res; + }, + src, dst); + } +} + +void CpuPoolingKernel::pooling3_f32_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) +{ + Iterator src(_src, window_src); + Iterator dst(_dst, window); + + constexpr const int pool_size = 3; + const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); + const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); + + const uint8_t *const src_top_ptr = _src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top))); + const uint8_t *const src_middle_ptr = _src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 1)); + const uint8_t *const src_bottom_ptr = _src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + 2)); + + execute_window_loop(window, [&](const Coordinates & id) + { + float32x4_t top_data = vld1q_f32(reinterpret_cast(src_top_ptr + src.offset())); + float32x4_t middle_data = vld1q_f32(reinterpret_cast(src_middle_ptr + src.offset())); + float32x4_t bottom_data = vld1q_f32(reinterpret_cast(src_bottom_ptr + src.offset())); + float32x2_t res = {}; + float final_res = 0; + + // Get power of 2 in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + top_data = vmulq_f32(top_data, top_data); + middle_data = vmulq_f32(middle_data, middle_data); + bottom_data = vmulq_f32(bottom_data, bottom_data); + } + + if(pooling_type != PoolingType::MAX) + { + // Calculate scale + float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + const float32x2_t scale_v = vdup_n_f32(scale); + + // Perform pooling + const float32x4_t sum_data = vaddq_f32(vaddq_f32(top_data, bottom_data), middle_data); + res = vpadd_f32(vget_high_f32(vsetq_lane_f32(0.f, sum_data, 3)), vget_low_f32(sum_data)); + res = vmul_f32(vpadd_f32(res, res), scale_v); + } + else + { + const float32x4_t max_data = vmaxq_f32(vmaxq_f32(top_data, bottom_data), middle_data); + res = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits::max(), max_data, 3)), vget_low_f32(max_data)); + res = vpmax_f32(res, res); + } + final_res = vget_lane_f32(res, 0); + + // Calculate square-root in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + final_res = sqrt(final_res); + } + + // Store result + *(reinterpret_cast(dst.ptr())) = final_res; + }, + src, dst); +} + +void CpuPoolingKernel::pooling7_f32_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) +{ + Iterator src(_src, window_src); + Iterator dst(_dst, window); + + constexpr const int pool_size = 7; + const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); + const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); + + std::array src_ptrs{ {} }; + for(int i = 0; i < pool_size; ++i) + { + src_ptrs[i] = _src->ptr_to_element(Coordinates(-static_cast(pool_pad_left), -static_cast(pool_pad_top) + i)); + } + + execute_window_loop(window, [&](const Coordinates & id) + { + float32x2_t res = {}; + float final_res = 0.f; + if(pooling_type != PoolingType::MAX) + { + // Calculate scale + float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + const float32x2_t scale_v = vdup_n_f32(scale); + + // Perform pooling + float32x4x2_t data = vld2q_f32(reinterpret_cast(src_ptrs[0] + src.offset())); + // Get power of 2 in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + data.val[0] = vmulq_f32(data.val[0], data.val[0]); + data.val[1] = vmulq_f32(data.val[1], data.val[1]); + } + float32x4_t sum_data = vaddq_f32(data.val[0], vsetq_lane_f32(0.f, data.val[1], 3)); + for(int i = 1; i < pool_size; ++i) + { + data = vld2q_f32(reinterpret_cast(src_ptrs[i] + src.offset())); + // Get power of 2 in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + data.val[0] = vmulq_f32(data.val[0], data.val[0]); + data.val[1] = vmulq_f32(data.val[1], data.val[1]); + } + sum_data = vaddq_f32(sum_data, data.val[0]); + sum_data = vaddq_f32(sum_data, vsetq_lane_f32(0.f, data.val[1], 3)); + } + res = vpadd_f32(vget_high_f32(sum_data), vget_low_f32(sum_data)); + res = vmul_f32(vpadd_f32(res, res), scale_v); + } + else + { + float32x4x2_t max_data = vld2q_f32(reinterpret_cast(src_ptrs[0] + src.offset())); + for(int i = 1; i < pool_size; ++i) + { + const float32x4x2_t data = vld2q_f32(reinterpret_cast(src_ptrs[i] + src.offset())); + max_data = vmax2q_f32(max_data, data); + } + res = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits::max(), max_data.val[1], 3)), vget_low_f32(max_data.val[1])); + res = vpmax_f32(res, vpmax_f32(vget_high_f32(max_data.val[0]), vget_low_f32(max_data.val[0]))); + res = vpmax_f32(res, res); + } + final_res = vget_lane_f32(res, 0); + + // Calculate square-root in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + final_res = sqrt(final_res); + } + + // Store result + *(reinterpret_cast(dst.ptr())) = final_res; + }, + src, dst); +} + +void CpuPoolingKernel::poolingMxN_f32_nhwc(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) +{ + if(_pool_info.pool_size == Size2D(2, 2) && pooling_type == PoolingType::MAX && _indices) + { + pooling2_f32_nhwc_maxpool_indices(window_src, window); + } + else + { + const int window_start_x = window.x().start(); + const int window_end_x = window.x().end(); + const int window_step_x = 4; + + Window window_out = window; + window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator src(_src, window_src); + Iterator dst(_dst, window_out); + + const int pool_size_x = _pool_info.is_global_pooling ? _src->info()->tensor_shape().y() : _pool_info.pool_size.width; + const int pool_size_y = _pool_info.is_global_pooling ? _src->info()->tensor_shape().z() : _pool_info.pool_size.height; + const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); + const int upper_bound_w = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = _src->info()->dimension(2) + (exclude_padding ? 0 : pool_pad_bottom); + + float32x4_t vres; + + execute_window_loop(window_out, [&](const Coordinates & id) + { + const int idx_width = id.y() * pool_stride_x; + const int idx_height = id.z() * pool_stride_y; + const int pool_limit_y = pool_pad_top - idx_height; + const int pool_limit_x = pool_pad_left - idx_width; + + const int pool_start_y = std::max(0, window_src.z().start() + pool_limit_y); + const int pool_end_y = std::min(pool_size_y, window_src.z().end() + pool_limit_y); + const int pool_start_x = std::max(0, window_src.y().start() + pool_limit_x); + const int pool_end_x = std::min(pool_size_x, window_src.y().end() + pool_limit_x); + + int x_off = window_start_x; + for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) + { + if(pooling_type != PoolingType::MAX) + { + // Calculate scale + const float scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, + pool_stride_y); + const float32x4_t scale_v = vdupq_n_f32(scale); + + // Perform pooling + vres = vdupq_n_f32(0.0f); + + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const float32x4_t data = vld1q_f32(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z())) + x_off); + + // Get power of 2 in case of l2 pooling and accumulate + if(pooling_type == PoolingType::L2) + { + vres = vmlaq_f32(vres, data, data); + } + else + { + vres = vaddq_f32(vres, data); + } + } + } + // Divide by scale + vres = vmulq_f32(vres, scale_v); + } + else + { + vres = vdupq_n_f32(std::numeric_limits::lowest()); + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const float32x4_t data = vld1q_f32(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z())) + x_off); + vres = vmaxq_f32(vres, data); + } + } + } + + // Calculate square-root in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + float32x4_t l2_res = { static_cast(sqrt(vgetq_lane_f32(vres, 0))), + static_cast(sqrt(vgetq_lane_f32(vres, 1))), + static_cast(sqrt(vgetq_lane_f32(vres, 2))), + static_cast(sqrt(vgetq_lane_f32(vres, 3))) + }; + vres = l2_res; + } + + // Store result + vst1q_f32(reinterpret_cast(dst.ptr()) + x_off, vres); + } + + // Left-overs loop + for(; x_off < window_end_x; ++x_off) + { + float res = 0.0f; + + if(pooling_type != PoolingType::MAX) + { + // Calculate scale + const float scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, + pool_stride_y); + + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const float data = *(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z())) + x_off); + + // Get power of 2 in case of l2 pooling and accumulate + if(pooling_type == PoolingType::L2) + { + res += data * data; + } + else + { + res += data; + } + } + } + + // Divide by scale + res *= scale; + } + else + { + res = std::numeric_limits::lowest(); + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const float data = *(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z())) + x_off); + res = std::max(res, data); + } + } + } + + // Calculate square-root in case of l2 pooling + if(pooling_type == PoolingType::L2) + { + res = std::sqrt(res); + } + + // Store result + *(reinterpret_cast(dst.ptr()) + x_off) = res; + } + }, + src, dst); + } +} + +void CpuPoolingKernel::pooling2_f32_nhwc_maxpool_indices(const Window &window_src, const Window &window) +{ + const int window_start_x = window.x().start(); + const int window_end_x = window.x().end(); + const int window_step_x = 4; + + Window window_out = window; + window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator src(_src, window_src); + Iterator dst(_dst, window_out); + Iterator indices(_indices, window_out); + + const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); + + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); + + float32x4_t vres; + float res; + + const int pad_right = _src->info()->padding().right; + const int in_stride_y = static_cast(_src->info()->strides_in_bytes().y()); + const int in_stride_z = static_cast(_src->info()->strides_in_bytes().z()); + + execute_window_loop(window_out, [&](const Coordinates & id) + { + const int idx_width = id.y() * pool_stride_x; + const int idx_height = id.z() * pool_stride_y; + const int pool_limit_y = pool_pad_top - idx_height; + const int pool_limit_x = pool_pad_left - idx_width; + + const int pool_start_y = std::max(0, window_src.z().start() + pool_limit_y); + const int pool_start_x = std::max(0, window_src.y().start() + pool_limit_x); + + const int in_x0_offset = (pool_start_x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (pool_start_y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z()); + const int in_x1_offset = (pool_start_x + 1 - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (pool_start_y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z()); + const int in_x2_offset = (pool_start_x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (pool_start_y + 1 - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z()); + const int in_x3_offset = (pool_start_x + 1 - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (pool_start_y + 1 - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z()); + + int x_off = window_start_x; + for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) + { + const auto in_x0_ptr = reinterpret_cast(src.ptr() + in_x0_offset); + const auto in_x1_ptr = reinterpret_cast(src.ptr() + in_x1_offset); + const auto in_x2_ptr = reinterpret_cast(src.ptr() + in_x2_offset); + const auto in_x3_ptr = reinterpret_cast(src.ptr() + in_x3_offset); + const auto v_x0 = vld1q_f32(in_x0_ptr + x_off); + const auto v_x1 = vld1q_f32(in_x1_ptr + x_off); + const auto v_x2 = vld1q_f32(in_x2_ptr + x_off); + const auto v_x3 = vld1q_f32(in_x3_ptr + x_off); + vres = vmaxq_f32(vmaxq_f32(v_x2, v_x3), vmaxq_f32(v_x0, v_x1)); + // Store result + vst1q_f32(reinterpret_cast(dst.ptr()) + x_off, vres); + + const uint32_t offset_base = offset_no_padding(src.offset(), id, *_src->info(), pool_stride_x, pool_stride_y); + const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float) + x_off; + const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float) - pad_right; + const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float) - pad_right * _src->info()->tensor_shape()[1]; + const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float) - pad_right; + const uint32x4_t voffset_x0 = { offset_x0, offset_x0 + 1, offset_x0 + 2, offset_x0 + 3 }; + const uint32x4_t voffset_x1 = { offset_x1, offset_x1 + 1, offset_x1 + 2, offset_x1 + 3 }; + const uint32x4_t voffset_x2 = { offset_x2, offset_x2 + 1, offset_x2 + 2, offset_x2 + 3 }; + const uint32x4_t voffset_x3 = { offset_x3, offset_x3 + 1, offset_x3 + 2, offset_x3 + 3 }; + const uint32x4_t tmp_indices0 = vbslq_u32(vcgeq_f32(v_x0, v_x1), voffset_x0, voffset_x1); + const uint32x4_t tmp_indices1 = vbslq_u32(vcgeq_f32(v_x2, v_x3), voffset_x2, voffset_x3); + const uint32x4_t tmp_indices2 = vbslq_u32(vcgeq_f32(vmaxq_f32(v_x0, v_x1), vmaxq_f32(v_x2, v_x3)), tmp_indices0, tmp_indices1); + + // Store indices + vst1q_u32(reinterpret_cast(indices.ptr()) + x_off, tmp_indices2); + } + + // Left-overs loop + for(; x_off < window_end_x; ++x_off) + { + const auto x0 = *(reinterpret_cast(src.ptr() + in_x0_offset) + x_off); + const auto x1 = *(reinterpret_cast(src.ptr() + in_x1_offset) + x_off); + const auto x2 = *(reinterpret_cast(src.ptr() + in_x2_offset) + x_off); + const auto x3 = *(reinterpret_cast(src.ptr() + in_x3_offset) + x_off); + res = std::max(std::max(x2, x3), std::max(x0, x1)); + + // Store result + *(reinterpret_cast(dst.ptr()) + x_off) = res; + + const uint32_t offset_base = offset_no_padding(src.offset(), id, *_src->info(), pool_stride_x, pool_stride_y); + const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float) + x_off; + const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float) - pad_right; + const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float) - pad_right * _src->info()->tensor_shape()[1]; + const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float) - pad_right; + const uint32_t tmp_idx0 = (x0 >= x1) ? offset_x0 : offset_x1; + const uint32_t tmp_idx1 = (x2 >= x3) ? offset_x2 : offset_x3; + const uint32_t tmp_idx2 = (std::max(x0, x1) >= std::max(x2, x3)) ? tmp_idx0 : tmp_idx1; + + // Store indices + *(reinterpret_cast(indices.ptr()) + x_off) = tmp_idx2; + } + }, + src, dst, indices); +} + +template +void CpuPoolingKernel::poolingMxN_q8_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) +{ + Iterator src(_src, window_src); + Iterator dst(_dst, window); + + /** NEON vector types */ + using q8x8_t = typename wrapper::traits::neon_vector::type; + using q16_t = typename wrapper::traits::promote_t; + using q16x8_t = typename wrapper::traits::neon_vector::type; + using q32_t = typename wrapper::traits::promote_t; + using q32x4_t = typename wrapper::traits::neon_vector::type; + + const int pool_size_x = _pool_info.is_global_pooling ? _src->info()->tensor_shape().x() : _pool_info.pool_size.width; + const int pool_size_y = _pool_info.is_global_pooling ? _src->info()->tensor_shape().y() : _pool_info.pool_size.height; + const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); + const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); + + const UniformQuantizationInfo &src_qinfo = _src->info()->quantization_info().uniform(); + const UniformQuantizationInfo &dst_qinfo = _dst->info()->quantization_info().uniform(); + + execute_window_loop(window, [&](const Coordinates & id) + { + T res = std::numeric_limits::min(); + + if(pooling_type != PoolingType::MAX) + { + q32x4_t vres = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); + q32_t sres = 0; + + // Calculate scale + const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); + + // Perform pooling + for(int y = 0; y < pool_size_y; ++y) + { + int x = 0; + for(; x <= (pool_size_x - 8); x += 8) + { + const q8x8_t data = wrapper::vload(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().y()))); + + const q16x8_t data_q16 = wrapper::vmovl(data); + vres = wrapper::vadd(vres, wrapper::vaddl(wrapper::vgethigh(data_q16), wrapper::vgetlow(data_q16))); + } + + // Leftover for loop + for(; x < pool_size_x; ++x) + { + T data = *(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().y()))); + sres += data; + } + } + + // Reduction + const auto tmp = wrapper::vpadd(wrapper::vgethigh(vres), wrapper::vgetlow(vres)); + sres += wrapper::vgetlane(tmp, 0) + wrapper::vgetlane(tmp, 1); + + // Divide by scale + res = static_cast(support::cpp11::round(sres * scale)); + } + else + { + q8x8_t vres = wrapper::vdup_n(std::numeric_limits::min(), wrapper::traits::vector_64_tag{}); + + for(int y = 0; y < pool_size_y; ++y) + { + int x = 0; + for(; x <= (pool_size_x - 8); x += 8) + { + const q8x8_t data = wrapper::vload(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().y()))); + vres = wrapper::vmax(vres, data); + } + // Leftover for loop + for(; x < pool_size_x; ++x) + { + const T data = *(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().y()))); + res = std::max(res, data); + } + } + + // Reduce max + vres = wrapper::vpmax(vres, vres); + vres = wrapper::vpmax(vres, vres); + vres = wrapper::vpmax(vres, vres); + + // Get max value + res = std::max(res, wrapper::vgetlane(vres, 0)); + } + // Store result + res = (src_qinfo != dst_qinfo) ? Qasymm8QuantizationHelper::quantize(Qasymm8QuantizationHelper::dequantize(res, src_qinfo), dst_qinfo) : res; + *(reinterpret_cast(dst.ptr())) = res; + }, + src, dst); +} + +template +void CpuPoolingKernel::poolingMxN_q8_nhwc(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) +{ + const int window_start_x = window.x().start(); + const int window_end_x = window.x().end(); + const int window_step_x = 16; + const int window_half_step_x = window_step_x / 2; + + Window window_out = window; + window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator src(_src, window_src); + Iterator dst(_dst, window_out); + + using q8x8_t = typename wrapper::traits::neon_vector::type; + using q8x16_t = typename wrapper::traits::neon_vector::type; + using q16_t = typename wrapper::traits::promote_t; + using q16x8_t = typename wrapper::traits::neon_vector::type; + using q32_t = typename wrapper::traits::promote_t; + using q32x4_t = typename wrapper::traits::neon_vector::type; + + const int pool_size_x = _pool_info.is_global_pooling ? _src->info()->tensor_shape().y() : _pool_info.pool_size.width; + const int pool_size_y = _pool_info.is_global_pooling ? _src->info()->tensor_shape().z() : _pool_info.pool_size.height; + const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); + const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); + const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); + const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); + + int pool_stride_x = 0; + int pool_stride_y = 0; + std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); + const int upper_bound_w = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_right); + const int upper_bound_h = _src->info()->dimension(2) + (exclude_padding ? 0 : pool_pad_bottom); + + const float32x4_t half_scale_v = vdupq_n_f32(0.5f); + const UniformQuantizationInfo src_qinfo = _src->info()->quantization_info().uniform(); + const UniformQuantizationInfo dst_qinfo = _dst->info()->quantization_info().uniform(); + + const float quant_rescale = dst_qinfo.scale / src_qinfo.scale; + // "new_offset" doesn't have to consider the "half_scale_v" in its computation + // With a requantization performed in a single step there won't be uncertainties introduced + const int32_t new_offset = dst_qinfo.offset - static_cast(static_cast(src_qinfo.offset) / quant_rescale); + + const float requant_scale = dst_qinfo.scale / src_qinfo.scale; + const int32_t requant_offset = dst_qinfo.offset - static_cast(static_cast(src_qinfo.offset) / requant_scale); + const UniformQuantizationInfo requant_qinfo = UniformQuantizationInfo(requant_scale, requant_offset); + + execute_window_loop(window_out, [&](const Coordinates & id) + { + const int idx_width = id.y() * pool_stride_x; + const int idx_height = id.z() * pool_stride_y; + const int pool_limit_y = pool_pad_top - idx_height; + const int pool_limit_x = pool_pad_left - idx_width; + + const int pool_start_y = std::max(0, window_src.z().start() + pool_limit_y); + const int pool_end_y = std::min(pool_size_y, window_src.z().end() + pool_limit_y); + const int pool_start_x = std::max(0, window_src.y().start() + pool_limit_x); + const int pool_end_x = std::min(pool_size_x, window_src.y().end() + pool_limit_x); + + int x_off = window_start_x; + for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) + { + if(pooling_type != PoolingType::MAX) + { + q32x4_t vres1 = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); + q32x4_t vres2 = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); + q32x4_t vres3 = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); + q32x4_t vres4 = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); + + // Calculate scale + const float scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, + pool_stride_y); + + // Perform pooling + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const q8x16_t data = wrapper::vloadq(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z())) + x_off); + + const q16x8_t data_q16 = wrapper::vmovl(wrapper::vgetlow(data)); + const q16x8_t data2_q16 = wrapper::vmovl(wrapper::vgethigh(data)); + vres1 = wrapper::vadd(vres1, wrapper::vmovl(wrapper::vgetlow(data_q16))); + vres2 = wrapper::vadd(vres2, wrapper::vmovl(wrapper::vgethigh(data_q16))); + vres3 = wrapper::vadd(vres3, wrapper::vmovl(wrapper::vgetlow(data2_q16))); + vres4 = wrapper::vadd(vres4, wrapper::vmovl(wrapper::vgethigh(data2_q16))); + } + } + + if(src_qinfo != dst_qinfo) + { + const float32x4x4_t vres = + { + { + vcvtq_f32_q32(vres1), + vcvtq_f32_q32(vres2), + vcvtq_f32_q32(vres3), + vcvtq_f32_q32(vres4), + } + }; + const auto requantized_dst = vrequantize_pooling_with_scale(vres, quant_rescale, scale, new_offset); + // Store result + wrapper::vstore(reinterpret_cast(dst.ptr()) + x_off, wrapper::vgetlow(requantized_dst)); + wrapper::vstore(reinterpret_cast(dst.ptr()) + x_off + 8, wrapper::vgethigh(requantized_dst)); + } + else + { + const float32x4_t scale_v = vdupq_n_f32(scale); + // Divide by scale and add 0.5f to round to nearest instead of rounding towards zero + vres1 = vcvtq_q32_f32(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres1), scale_v)); + vres2 = vcvtq_q32_f32(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres2), scale_v)); + vres3 = vcvtq_q32_f32(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres3), scale_v)); + vres4 = vcvtq_q32_f32(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres4), scale_v)); + + const q8x8_t res1 = wrapper::vmovn(wrapper::vcombine(wrapper::vmovn(vres1), wrapper::vmovn(vres2))); + const q8x8_t res2 = wrapper::vmovn(wrapper::vcombine(wrapper::vmovn(vres3), wrapper::vmovn(vres4))); + // Store result + wrapper::vstore(reinterpret_cast(dst.ptr()) + x_off, res1); + wrapper::vstore(reinterpret_cast(dst.ptr()) + x_off + 8, res2); + } + } + else + { + q8x16_t vres = wrapper::vdup_n(std::numeric_limits::min(), wrapper::traits::vector_128_tag{}); + + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const q8x16_t data = wrapper::vloadq(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z())) + x_off); + vres = wrapper::vmax(vres, data); + } + } + + // Store result + wrapper::vstore(reinterpret_cast(dst.ptr()) + x_off, (src_qinfo != dst_qinfo) ? vrequantize_pooling(wrapper::vgetlow(vres), wrapper::vgethigh(vres), + requant_qinfo) : + vres); + } + } + + if(pooling_type == PoolingType::MAX) + { + for(; x_off <= (window_end_x - window_half_step_x); x_off += window_half_step_x) + { + q8x8_t vres = wrapper::vdup_n(std::numeric_limits::min(), wrapper::traits::vector_64_tag{}); + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const q8x8_t data = wrapper::vload(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z())) + x_off); + vres = wrapper::vmax(vres, data); + } + } + + // Store result + wrapper::vstore(reinterpret_cast(dst.ptr()) + x_off, + (src_qinfo != dst_qinfo) ? vrequantize_pooling(vres, requant_qinfo) : vres); + } + } + + // Left-overs loop + for(; x_off < window_end_x; ++x_off) + { + if(pooling_type != PoolingType::MAX) + { + q32_t res = static_cast(0.f); + + // Calculate scale + const float scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, + pool_stride_y); + + // Perform pooling + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const T data = *(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z())) + x_off); + res += data; + } + } + + if(src_qinfo != dst_qinfo) + { + const float res_f = static_cast(res); + const float new_scale = quant_rescale / scale; + const auto requantized_dst = quantize(res_f, UniformQuantizationInfo(new_scale, new_offset)); + + // Store result + *(reinterpret_cast(dst.ptr()) + x_off) = requantized_dst; + } + else + { + // Divide by scale and add 0.5f to round to nearest instead of rounding towards zero + res = static_cast(0.5f + static_cast(res) * scale); + + // Store result + *(reinterpret_cast(dst.ptr()) + x_off) = res; + } + } + else + { + T res = std::numeric_limits::min(); + + for(int y = pool_start_y; y < pool_end_y; ++y) + { + for(int x = pool_start_x; x < pool_end_x; ++x) + { + const T data = *(reinterpret_cast(src.ptr() + (x - pool_pad_left) * static_cast(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast + (_src->info()->strides_in_bytes().z())) + x_off); + res = std::max(res, data); + } + } + + // Store result + if(src_qinfo != dst_qinfo) + { + const float res_f = static_cast(res); + *(reinterpret_cast(dst.ptr()) + x_off) = quantize(res_f, requant_qinfo); + } + else + { + *(reinterpret_cast(dst.ptr()) + x_off) = res; + } + } + } + + }, + src, dst); +} + +Status CpuPoolingKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &pool_info, const ITensorInfo *indices) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src); + + unsigned int pooled_w = 0; + unsigned int pooled_h = 0; + unsigned int num_elems_processed_per_iteration = 0; + BorderSize border_size(0); + + const bool is_global_pooling = pool_info.is_global_pooling; + unsigned int pool_size_x = 0; + unsigned int pool_size_y = 0; + + // Get data layout + const auto data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : pool_info.data_layout; + const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + + pool_size_x = is_global_pooling ? src->dimension(idx_width) : pool_info.pool_size.width; + pool_size_y = is_global_pooling ? src->dimension(idx_height) : pool_info.pool_size.height; + + // Validate pool info before calling scaled_dimensions + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_pool_info(pool_size_x, pool_size_y)); + + // Check dst dimensions + std::tie(pooled_w, pooled_h) = scaled_dimensions(src->dimension(idx_width), + src->dimension(idx_height), + pool_size_x, + pool_size_y, + pool_info.pad_stride_info); + + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, pool_info, pooled_w, pooled_h, indices, Size2D(pool_size_x, pool_size_y))); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), dst->clone().get(), + (indices) ? indices->clone().get() : nullptr, pool_info, num_elems_processed_per_iteration, border_size, pooled_w, pooled_h, + pool_size_x, pool_size_y) + .first); + + return Status{}; +} + +void CpuPoolingKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window); + ARM_COMPUTE_ERROR_ON(_func == nullptr); + + _src = tensors.get_const_tensor(TensorType::ACL_SRC_0); + _dst = tensors.get_tensor(TensorType::ACL_DST_0); + _indices = tensors.get_tensor(TensorType::ACL_DST_1); + + const unsigned int pool_stride_x = _pool_info.pad_stride_info.stride().first; + const unsigned int pool_stride_y = _pool_info.pad_stride_info.stride().second; + const unsigned int pool_size = _pool_info.pool_size.width; + const bool exclude_padding = _pool_info.exclude_padding; + + Window window_src(window); + if(_data_layout == DataLayout::NCHW) + { + // Set step for src in x and y direction for the src + unsigned int window_x_inc = 0; + switch(_src->info()->data_type()) + { + case DataType::QASYMM8: + case DataType::QASYMM8_SIGNED: + { + window_x_inc = pool_stride_x; + if((pool_size == 2 || pool_size == 3) && pool_stride_x < 3) + { + window_x_inc = (pool_stride_x == 2) ? _num_elems_processed_per_iteration * 2 : _num_elems_processed_per_iteration; + } + break; + } + + case DataType::F16: + case DataType::F32: + { + window_x_inc = pool_stride_x; + break; + } + default: + { + ARM_COMPUTE_ERROR("Not supported"); + } + } + window_src.set(Window::DimX, Window::Dimension(window.x().start() * pool_stride_x, window.x().end() * pool_stride_x, window_x_inc)); + window_src.set(Window::DimY, Window::Dimension(window.y().start() * pool_stride_y, window.y().end() * pool_stride_y, pool_stride_y)); + } + else + { + window_src.set(Window::DimX, Window::Dimension(0, 1, 1)); + window_src.set(Window::DimY, Window::Dimension(0, _src->info()->dimension(1), pool_stride_x)); + window_src.set(Window::DimZ, Window::Dimension(0, _src->info()->dimension(2), pool_stride_y)); + } + + // Run function + (this->*_func)(window_src, window, _pool_info.pool_type, exclude_padding); +} +} // namespace kernels +} // namespace cpu +} // namespace arm_compute diff --git a/src/core/cpu/kernels/CpuPoolingKernel.h b/src/core/cpu/kernels/CpuPoolingKernel.h new file mode 100644 index 0000000000..036e43650e --- /dev/null +++ b/src/core/cpu/kernels/CpuPoolingKernel.h @@ -0,0 +1,226 @@ +/* + * Copyright (c) 2017-2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ARM_COMPUTE_CPU_POOLING_KERNEL_H +#define ARM_COMPUTE_CPU_POOLING_KERNEL_H + +#include "arm_compute/core/Types.h" +#include "src/core/common/Macros.h" +#include "src/core/cpu/ICpuKernel.h" + +namespace arm_compute +{ +namespace cpu +{ +namespace kernels +{ +/** Interface for the pooling layer kernel */ +class CpuPoolingKernel : public ICpuKernel +{ +public: + const char *name() const override + { + return "CpuPoolingKernel"; + } + /** Default constructor */ + CpuPoolingKernel() = default; + ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuPoolingKernel); + /** Configure kernel for a given list of arguments + * + * @note F16 are supported for pool sizes 2 and 3 only + * + * @param[in] src Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. + * @param[out] dst Destination tensor info. Data types supported: Same as @p src. + * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo. + * @param[out] indices (optional) The indices of the maximal values. Data type supported: U32. + */ + void configure(ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &pool_info, ITensorInfo *indices = nullptr); + /** Static function to check if given info will lead to a valid configuration of @ref CpuPoolingKernel + * + * @note F16 are supported for pool sizes 2 and 3 only + * + * @param[in] src Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. + * @param[in] dst Destination tensor info. Data types supported: Same as @p src. + * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo. + * @param[in] indices (optional) The indices of the maximal values. Data type supported: U32. + * + * @return a status + */ + static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &pool_info, const ITensorInfo *indices = nullptr); + + // Inherited methods overridden: + void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override; + BorderSize border_size() const override; + +private: + /** Function to perform 2x2 pooling. + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + * @param[in] pooling_type Pooling operation to be computed. + * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. + */ + void pooling2_f32_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding = false); + /** Function to perform 2x2 pooling and compute the pooling indices. The indices can be used for max unpool. + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + */ + void pooling2_f32_nhwc_maxpool_indices(const Window &window_src, const Window &window); + /** Function to perform MxN pooling for 32-bit floating point values. + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + * @param[in] pooling_type Pooling operation to be computed. + * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. + */ + void poolingMxN_f32_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding = false); + /** Function to perform MxN pooling for 32-bit floating point values (NHWC). + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + * @param[in] pooling_type Pooling operation to be computed. + * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. + */ + void poolingMxN_f32_nhwc(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding = false); + /** Function to perform 7x7 pooling. + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + * @param[in] pooling_type Pooling operation to be computed. + * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. + */ + void pooling7_f32_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding = false); + /** Function to perform 3x3 pooling. + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + * @param[in] pooling_type Pooling operation to be computed. + * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. + */ + void pooling3_f32_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding = false); + /** Function to perform 2x2 pooling for float16_t. + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + * @param[in] pooling_type Pooling operation to be computed. + * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. + */ + void pooling2_f16_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding = false); + /** Function to perform 2x2 pooling and compute the pooling indices for FP32/FP16. The indices can be used for max unpool. + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + */ + template + void pooling2_nchw_maxpool_indices(const Window &window_src, const Window &window); + /** Function to perform 2x2 pooling and compute the pooling indices. The indices can be used for max unpool. + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + */ + void pooling2_f16_nhwc_maxpool_indices(const Window &window_src, const Window &window); + /** Function to perform 3x3 pooling. + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + * @param[in] pooling_type Pooling operation to be computed. + * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. + */ + void pooling3_f16_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding = false); + /** Function to perform MxN pooling for 16-bit floating point values. + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + * @param[in] pooling_type Pooling operation to be computed. + * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. + */ + void poolingMxN_f16_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding = false); + /** Function to perform MxN pooling for 16-bit floating point values. (NHWC) + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + * @param[in] pooling_type Pooling operation to be computed. + * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. + */ + void poolingMxN_f16_nhwc(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding = false); + /** Template function to perform 2x2 pooling for 8bit quantized fixed point. (NCHW) + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + * @param[in] pooling_type Pooling operation to be computed. + * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. + */ + template + void pooling2_q8_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding = false); + /** Template function to perform 3x3 pooling for 8bit quantized fixed point. (NCHW) + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + * @param[in] pooling_type Pooling operation to be computed. + * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. + */ + template + void pooling3_q8_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding = false); + /** Template function to perform MxN pooling for 8-bit quantized. (NCHW) + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + * @param[in] pooling_type Pooling operation to be computed. + * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. + */ + template + void poolingMxN_q8_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding = false); + /** Template function to perform MxN pooling for 8-bit quantized. (NHWC) + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + * @param[in] pooling_type Pooling operation to be computed. + * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. + */ + template + void poolingMxN_q8_nhwc(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding = false); + /** Common signature for all the specialised Pooling functions + * + * @param[in] window_src src region on which to execute the kernel. + * @param[in] window dst region on which to execute the kernel. + * @param[in] pooling_type Pooling operation to be computed. + * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. + */ + using PoolingFunction = void (CpuPoolingKernel::*)(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding); + +private: + PoolingFunction _func{ nullptr }; + const ITensor *_src{ nullptr }; + ITensor *_dst{ nullptr }; + ITensor *_indices{ nullptr }; + PoolingLayerInfo _pool_info{}; + DataLayout _data_layout{ DataLayout::UNKNOWN }; + unsigned int _num_elems_processed_per_iteration{ 0 }; + BorderSize _border_size{ 0 }; + bool _is_square{ false }; +}; +} // namespace kernels +} // namespace cpu +} // namespace arm_compute +#endif /*ARM_COMPUTE_CPU_POOLING_KERNEL_H */ diff --git a/src/runtime/NEON/functions/NEPoolingAssemblyDispatch.cpp b/src/runtime/NEON/functions/NEPoolingAssemblyDispatch.cpp deleted file mode 100644 index 427cd2eb77..0000000000 --- a/src/runtime/NEON/functions/NEPoolingAssemblyDispatch.cpp +++ /dev/null @@ -1,156 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/runtime/NEON/functions/NEPoolingAssemblyDispatch.h" - -#include "arm_compute/core/ITensor.h" -#include "arm_compute/runtime/NEON/NEScheduler.h" -#include "src/core/CPP/Validate.h" -#include "src/core/NEON/kernels/assembly/NEPoolingAssemblyWrapperKernel.h" - -namespace arm_compute -{ -namespace experimental -{ -NEPoolingAssemblyDispatch::~NEPoolingAssemblyDispatch() = default; - -void NEPoolingAssemblyDispatch::configure(const ITensorInfo *input, ITensorInfo *output, const PoolingLayerInfo &info) -{ - const CPUInfo &ci = NEScheduler::get().cpu_info(); - const unsigned int num_threads = NEScheduler::get().num_threads(); - - // If we don't support a combination of data types, silently return: it is the caller's responsibility to check if configure() was successful via is_configured() - if(!NEPoolingAssemblyDispatch::validate(input, output, info)) - { - return; - } - - auto pooling_wrapper = std::make_unique(); - ARM_COMPUTE_ERROR_ON(pooling_wrapper == nullptr); - pooling_wrapper->configure(input, output, info, ci); - - // Check if we have Global Pooling Layer - _is_global_pooling_layer = (input->dimension(2) == info.pool_size.width) && (input->dimension(1) == info.pool_size.height); - - // Set workspace requirements - const unsigned int alignment = 4096; - _workspace.push_back(MemoryInfo(TensorType::ACL_DST_1, pooling_wrapper->get_working_size(num_threads), alignment)); - - _kernel = std::move(pooling_wrapper); -} - -Status NEPoolingAssemblyDispatch::validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &info) -{ - return NEPoolingAssemblyWrapperKernel::validate(input, output, info); -} - -bool NEPoolingAssemblyDispatch::is_configured() const -{ - return _kernel != nullptr; -} - -void NEPoolingAssemblyDispatch::run(ITensorPack &tensors) -{ - if(tensors.empty()) - { - ARM_COMPUTE_ERROR("No inputs provided"); - } - - if(_is_global_pooling_layer) - { - NEScheduler::get().schedule_op(_kernel.get(), Window::DimX, _kernel->window(), tensors); - } - else - { - NEScheduler::get().schedule_op(_kernel.get(), Window::DimY, _kernel->window(), tensors); - } -} -} // namespace experimental - -struct NEPoolingAssemblyDispatch::Impl -{ - const ITensor *src{ nullptr }; - ITensor *dst{ nullptr }; - ITensor *workspace{ nullptr }; - std::unique_ptr op{ nullptr }; -}; - -NEPoolingAssemblyDispatch::NEPoolingAssemblyDispatch(NEPoolingAssemblyDispatch &&) = default; - -NEPoolingAssemblyDispatch &NEPoolingAssemblyDispatch::operator=(NEPoolingAssemblyDispatch &&) = default; - -NEPoolingAssemblyDispatch::~NEPoolingAssemblyDispatch() = default; - -NEPoolingAssemblyDispatch::NEPoolingAssemblyDispatch(std::shared_ptr memory_manager) - : _impl(std::make_unique()), - _memory_group(std::move(memory_manager)), - _workspace() -{ -} - -void NEPoolingAssemblyDispatch::configure(const ITensor *input, ITensor *output, const PoolingLayerInfo &info) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - - _impl->src = input; - _impl->dst = output; - _impl->workspace = &_workspace; - - _impl->op = std::make_unique(); - _impl->op->configure(input->info(), output->info(), info); - - const auto workspace = _impl->op->workspace().at(0); - if(workspace.size > 0) - { - // Allocate workspace - allocate_workspace(workspace.size, workspace.alignment); - } -} - -Status NEPoolingAssemblyDispatch::validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &info) -{ - return experimental::NEPoolingAssemblyDispatch::validate(input, output, info); -} - -bool NEPoolingAssemblyDispatch::is_configured() const -{ - return _impl->op->is_configured(); -} - -void NEPoolingAssemblyDispatch::run() -{ - ITensorPack pack; - pack.add_tensor(TensorType::ACL_SRC, _impl->src); - pack.add_tensor(TensorType::ACL_DST_0, _impl->dst); - pack.add_tensor(TensorType::ACL_DST_1, _impl->workspace); - _impl->op->run(pack); -} - -void NEPoolingAssemblyDispatch::allocate_workspace(size_t workspace_size, size_t alignment) -{ - ARM_COMPUTE_ERROR_ON_MSG(workspace_size == 0, "size cannot be 0"); - _workspace.allocator()->init(TensorInfo(TensorShape{ (workspace_size + alignment /* FIXME: remove alignment after COMPMID-1088 */) }, 1, DataType::S8), alignment); - _memory_group.manage(&_workspace); - _workspace.allocator()->allocate(); -} -} //namespace arm_compute diff --git a/src/runtime/NEON/functions/NEPoolingAssemblyDispatch.h b/src/runtime/NEON/functions/NEPoolingAssemblyDispatch.h deleted file mode 100644 index f6d232b931..0000000000 --- a/src/runtime/NEON/functions/NEPoolingAssemblyDispatch.h +++ /dev/null @@ -1,149 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_NEPOOLINGASSEMBLYDISPATCH_H -#define ARM_COMPUTE_NEPOOLINGASSEMBLYDISPATCH_H - -#include "arm_compute/runtime/IFunction.h" -#include "arm_compute/runtime/IMemoryManager.h" -#include "arm_compute/runtime/MemoryGroup.h" -#include "arm_compute/runtime/NEON/INEOperator.h" -#include "arm_compute/runtime/Tensor.h" -#include "src/core/NEON/INEKernel.h" - -namespace arm_compute -{ -// Forward Declarations -class ITensor; -struct PoolingLayerInfo; - -/** Assembly kernel glue */ -class NEPoolingAssemblyDispatch : public IFunction -{ -public: - /** Constructor */ - NEPoolingAssemblyDispatch(std::shared_ptr memory_manager = nullptr); - /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEPoolingAssemblyDispatch(const NEPoolingAssemblyDispatch &) = delete; - /** Default move constructor */ - NEPoolingAssemblyDispatch(NEPoolingAssemblyDispatch &&); - /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEPoolingAssemblyDispatch &operator=(const NEPoolingAssemblyDispatch &) = delete; - /** Default move assignment operator */ - NEPoolingAssemblyDispatch &operator=(NEPoolingAssemblyDispatch &&); - /** Destructor */ - ~NEPoolingAssemblyDispatch(); - - /** If supported create an assembly routine, else fallback to Compute Library function. - * - * @param[in] input Input tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. - * @param[out] output Output tensor to store the result of pooling. Data types supported: same as @p input. - * @param[in] info Pooling meta-data - */ - void configure(const ITensor *input, ITensor *output, const PoolingLayerInfo &info); - - /** Indicates whether or not this function can be used to process the given parameters. - * - * @param[in] input Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. - * @param[in] output Output tensor to store the result of pooling. Data types supported: same as @p input. - * @param[in] info Pooling meta-data - * - * @return a status. - */ - static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &info); - - /** Was the function successfully configured ? - * - * @return True if the function is configured and ready to run - */ - bool is_configured() const; - - // Inherited methods overridden: - void run() override; - -private: - /** Helper function to allocate memory for the workspace needed by the - * assembly kernels - * - * @param[in] workspace_size Total size of the workspace. - * @param[in] alignment Alignment requirement in bytes. - */ - void allocate_workspace(size_t workspace_size, size_t alignment); - - struct Impl; - std::unique_ptr _impl; - - MemoryGroup _memory_group{}; - Tensor _workspace{}; -}; - -namespace experimental -{ -/** Basic function to run pooling assembly kernels */ -class NEPoolingAssemblyDispatch : public INEOperator -{ -public: - /** Constructor */ - NEPoolingAssemblyDispatch() = default; - /** Prevent instances of this class from being copied */ - NEPoolingAssemblyDispatch(const NEPoolingAssemblyDispatch &) = delete; - /** Default move constructor */ - NEPoolingAssemblyDispatch(NEPoolingAssemblyDispatch &&) = default; - /** Prevent instances of this class from being copied */ - NEPoolingAssemblyDispatch &operator=(const NEPoolingAssemblyDispatch &) = delete; - /** Default move assignment operator */ - NEPoolingAssemblyDispatch &operator=(NEPoolingAssemblyDispatch &&) = default; - /** Destructor */ - ~NEPoolingAssemblyDispatch(); - - /** If supported create an assembly routine, else fallback to Compute Library function. - * - * @param[in] input Input tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. - * @param[out] output Output tensor to store the result of pooling. Data types supported: same as @p input. - * @param[in] info Pooling meta-data - */ - void configure(const ITensorInfo *input, ITensorInfo *output, const PoolingLayerInfo &info); - - /** Indicates whether or not this function can be used to process the given parameters. - * - * @param[in] input Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. - * @param[in] output Output tensor to store the result of pooling. Data types supported: same as @p input. - * @param[in] info Pooling meta-data - * - * @return a status. - */ - static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &info); - /** Was the function successfully configured ? - * - * @return True if the function is configured and ready to run - */ - bool is_configured() const; - // Run method overriden - void run(ITensorPack &tensors) override; - -private: - bool _is_global_pooling_layer{ false }; -}; -} // namespace experimental -} // namespace arm_compute -#endif /* ARM_COMPUTE_NEPOOLINGASSEMBLYDISPATCH_H */ diff --git a/src/runtime/NEON/functions/NEPoolingLayer.cpp b/src/runtime/NEON/functions/NEPoolingLayer.cpp index 0c857b54dc..dd7a3a337e 100644 --- a/src/runtime/NEON/functions/NEPoolingLayer.cpp +++ b/src/runtime/NEON/functions/NEPoolingLayer.cpp @@ -23,103 +23,48 @@ */ #include "arm_compute/runtime/NEON/functions/NEPoolingLayer.h" -#include "arm_compute/core/ITensor.h" -#include "arm_compute/runtime/NEON/NEScheduler.h" -#include "src/core/NEON/kernels/NEFillBorderKernel.h" -#include "src/core/NEON/kernels/NEPoolingLayerKernel.h" -#include "src/runtime/NEON/functions/NEPoolingAssemblyDispatch.h" +#include "arm_compute/core/Validate.h" +#include "src/runtime/cpu/operators/CpuPooling.h" namespace arm_compute { +struct NEPoolingLayer::Impl +{ + ITensor *src{ nullptr }; + ITensor *dst{ nullptr }; + ITensor *indices{ nullptr }; + std::shared_ptr memory_manager{ nullptr }; + std::unique_ptr op{ nullptr }; +}; + NEPoolingLayer::~NEPoolingLayer() = default; NEPoolingLayer::NEPoolingLayer(std::shared_ptr memory_manager) - : _memory_manager(std::move(memory_manager)), _pooling_layer_kernel(), _border_handler(), _asm_glue(), _is_global_pooling_layer(false), _data_layout(DataLayout::NCHW) + : _impl(std::make_unique()) { + _impl->memory_manager = std::move(memory_manager); } void NEPoolingLayer::configure(ITensor *input, ITensor *output, const PoolingLayerInfo &pool_info, ITensor *indices) { - // Check if we can run assembly kernels. Currently, indices are not supported by those kernels - const bool run_optimised = bool(NEPoolingAssemblyDispatch::validate(input->info(), output->info(), pool_info)) && (indices == nullptr); - - if(run_optimised) - { - _asm_glue = std::make_unique(_memory_manager); - _asm_glue->configure(input, output, pool_info); - ARM_COMPUTE_ERROR_ON(!_asm_glue->is_configured()); - } - else - { - // Check if we have Global Pooling Layer - _is_global_pooling_layer = (input->info()->dimension(0) == pool_info.pool_size.width) && (input->info()->dimension(1) == pool_info.pool_size.height); - - // Get data layout - _data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? input->info()->data_layout() : pool_info.data_layout; - - // Configure pooling kernel - _pooling_layer_kernel = std::make_unique(); - _pooling_layer_kernel->configure(input, output, pool_info, indices); - - switch(_data_layout) - { - case DataLayout::NCHW: - { - // Configure border depending on operation required (quantize border in case of asymmetric data_type) - BorderMode border_mode = (!indices && pool_info.pool_type == PoolingType::MAX) ? BorderMode::REPLICATE : BorderMode::CONSTANT; - PixelValue zero_value((indices) ? std::numeric_limits::min() : 0.f); - if(is_data_type_quantized_asymmetric(input->info()->data_type()) && !pool_info.exclude_padding) - { - zero_value = PixelValue(0, input->info()->data_type(), input->info()->quantization_info()); - } - _border_handler = std::make_unique(); - _border_handler->configure(input, _pooling_layer_kernel->border_size(), border_mode, zero_value); - break; - } - case DataLayout::NHWC: - break; - default: - ARM_COMPUTE_ERROR("Data layout not supported"); - } - } + _impl->src = input; + _impl->dst = output; + _impl->indices = indices; + _impl->op = std::make_unique(_impl->memory_manager); + _impl->op->configure(input->info(), output->info(), pool_info, (indices) ? indices->info() : nullptr); } Status NEPoolingLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, const ITensorInfo *indices) { - const bool run_optimised = bool(NEPoolingAssemblyDispatch::validate(input, output, pool_info)) && (indices == nullptr); - - if(run_optimised) - { - return Status{}; - } - - return NEPoolingLayerKernel::validate(input, output, pool_info, indices); + return cpu::CpuPooling::validate(input, output, pool_info, indices); } void NEPoolingLayer::run() { - if(_asm_glue && _asm_glue->is_configured()) - { - _asm_glue->run(); - } - else - { - switch(_data_layout) - { - case DataLayout::NCHW: - // Fill border - NEScheduler::get().schedule(_border_handler.get(), Window::DimY); - - // Run pooling layer - NEScheduler::get().schedule(_pooling_layer_kernel.get(), _is_global_pooling_layer ? Window::DimZ : Window::DimY); - break; - case DataLayout::NHWC: - // Run pooling layer - NEScheduler::get().schedule(_pooling_layer_kernel.get(), Window::DimX); - break; - default: - ARM_COMPUTE_ERROR("Data layout not supported"); - } - } + ITensorPack pack; + pack.add_tensor(TensorType::ACL_SRC, _impl->src); + pack.add_tensor(TensorType::ACL_DST_0, _impl->dst); + pack.add_tensor(TensorType::ACL_DST_1, _impl->indices); + _impl->op->run(pack); } } // namespace arm_compute diff --git a/src/runtime/cpu/operators/CpuPooling.cpp b/src/runtime/cpu/operators/CpuPooling.cpp new file mode 100644 index 0000000000..0b9b38d079 --- /dev/null +++ b/src/runtime/cpu/operators/CpuPooling.cpp @@ -0,0 +1,130 @@ +/* + * Copyright (c) 2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "src/runtime/cpu/operators/CpuPooling.h" + +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/runtime/NEON/NEScheduler.h" +#include "src/core/NEON/kernels/NEFillBorderKernel.h" +#include "src/core/cpu/kernels/CpuPoolingKernel.h" + +namespace arm_compute +{ +namespace cpu +{ +CpuPooling::CpuPooling(std::shared_ptr memory_manager) + : _memory_manager(std::move(memory_manager)), _pooling_layer_kernel(), _border_handler(), _asm_glue(), _is_global_pooling_layer(false), _data_layout(DataLayout::NCHW) +{ +} + +CpuPooling::~CpuPooling() = default; + +void CpuPooling::configure(ITensorInfo *input, ITensorInfo *output, const PoolingLayerInfo &pool_info, ITensorInfo *indices) +{ + // Check if we can run assembly kernels. Currently, indices are not supported by those kernels + const bool run_optimised = bool(CpuPoolingAssemblyDispatch::validate(input, output, pool_info)) && (indices == nullptr); + + if(run_optimised) + { + _asm_glue = std::make_unique(_memory_manager); + _asm_glue->configure(input, output, pool_info); + ARM_COMPUTE_ERROR_ON(!_asm_glue->is_configured()); + } + else + { + // Check if we have Global Pooling Layer + _is_global_pooling_layer = (input->dimension(0) == pool_info.pool_size.width) && (input->dimension(1) == pool_info.pool_size.height); + + // Get data layout + _data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? input->data_layout() : pool_info.data_layout; + + // Configure pooling kernel + auto k = std::make_unique(); + k->configure(input, output, pool_info, indices); + _pooling_layer_kernel = std::move(k); + + switch(_data_layout) + { + case DataLayout::NCHW: + { + // Configure border depending on operation required (quantize border in case of asymmetric data_type) + BorderMode border_mode = (!indices && pool_info.pool_type == PoolingType::MAX) ? BorderMode::REPLICATE : BorderMode::CONSTANT; + PixelValue zero_value((indices) ? std::numeric_limits::min() : 0.f); + if(is_data_type_quantized_asymmetric(input->data_type()) && !pool_info.exclude_padding) + { + zero_value = PixelValue(0, input->data_type(), input->quantization_info()); + } + auto b = std::make_unique(); + b->configure(input, _pooling_layer_kernel->border_size(), border_mode, zero_value); + _border_handler = std::move(b); + break; + } + case DataLayout::NHWC: + break; + default: + ARM_COMPUTE_ERROR("Data layout not supported"); + } + } +} + +Status CpuPooling::validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, const ITensorInfo *indices) +{ + const bool run_optimised = bool(CpuPoolingAssemblyDispatch::validate(input, output, pool_info)) && (indices == nullptr); + + if(run_optimised) + { + return Status{}; + } + + return kernels::CpuPoolingKernel::validate(input, output, pool_info, indices); +} + +void CpuPooling::run(ITensorPack &tensors) +{ + if(_asm_glue && _asm_glue->is_configured()) + { + _asm_glue->run(tensors); + } + else + { + switch(_data_layout) + { + case DataLayout::NCHW: + // Fill border + NEScheduler::get().schedule_op(_border_handler.get(), Window::DimY, _border_handler->window(), tensors); + + // Run pooling layer + NEScheduler::get().schedule_op(_pooling_layer_kernel.get(), _is_global_pooling_layer ? Window::DimZ : Window::DimY, _pooling_layer_kernel->window(), tensors); + break; + case DataLayout::NHWC: + // Run pooling layer + NEScheduler::get().schedule_op(_pooling_layer_kernel.get(), Window::DimX, _pooling_layer_kernel->window(), tensors); + break; + default: + ARM_COMPUTE_ERROR("Data layout not supported"); + } + } +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/runtime/cpu/operators/CpuPooling.h b/src/runtime/cpu/operators/CpuPooling.h new file mode 100644 index 0000000000..aa607b4b44 --- /dev/null +++ b/src/runtime/cpu/operators/CpuPooling.h @@ -0,0 +1,102 @@ +/* + * Copyright (c) 2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ARM_COMPUTE_CPU_POOLING_H +#define ARM_COMPUTE_CPU_POOLING_H + +#include "src/runtime/cpu/ICpuOperator.h" + +#include "arm_compute/runtime/IMemoryManager.h" +#include "src/runtime/cpu/operators/CpuPoolingAssemblyDispatch.h" + +#include + +namespace arm_compute +{ +// Forward Declarations +struct PoolingLayerInfo; + +namespace cpu +{ +// Forward Declarations +class CpuPoolingAssemblyDispatch; +/** Basic function to simulate a pooling layer with the specified pooling operation. This function calls the following NEON kernels: + * + * -# @ref NEFillBorderKernel (executed if padding size is different from zero) + * -# @ref kernels::CpuPoolingKernel + * -# @ref CpuPoolingAssemblyDispatch + */ +class CpuPooling : public ICpuOperator +{ +public: + /** Constructor */ + CpuPooling(std::shared_ptr memory_manager = nullptr); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CpuPooling(const CpuPooling &) = delete; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CpuPooling &operator=(const CpuPooling &) = delete; + /** Prevent instances of this class from being moved (As this class contains non movable objects) */ + CpuPooling(CpuPooling &&) = delete; + /** Prevent instances of this class from being moved (As this class contains non movable objects) */ + CpuPooling &operator=(CpuPooling &&) = delete; + /** Default destructor */ + ~CpuPooling(); + /** Set the src and dst tensors. + * + * @note F16 is supported for pool sizes 2 and 3 only + * + * @param[in, out] src Source tensor info. (Written to only when padding != 0) Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. + * @param[out] dst Destination tensor info. Data types supported: same as @p src. + * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo. + * @param[out] indices (optional) The indices of the maximal values. Data type supported: U32. + */ + void configure(ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &pool_info, ITensorInfo *indices = nullptr); + /** Static function to check if given info will lead to a valid configuration of @ref CpuPooling + * + * @note F16 is supported for pool sizes 2 and 3 only + * + * @param[in] src Source tensor info. (Written to only when padding != 0) Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. + * @param[in] dst Destination tensor info. Data types supported: same as @p src. + * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo. + * @param[in] indices (optional) Tensor info of the indices of the maximal values. Data type supported: U32. + * + * @return a status + */ + static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &pool_info, const ITensorInfo *indices = nullptr); + + // Inherited methods overridden: + void run(ITensorPack &tensors) override; + +private: + std::shared_ptr _memory_manager; + + std::unique_ptr _pooling_layer_kernel; + std::unique_ptr _border_handler; + std::unique_ptr _asm_glue; + + bool _is_global_pooling_layer; + DataLayout _data_layout; +}; +} // namespace cpu +} // namespace arm_compute +#endif /* ARM_COMPUTE_CPU_POOLING_H */ diff --git a/src/runtime/cpu/operators/CpuPoolingAssemblyDispatch.cpp b/src/runtime/cpu/operators/CpuPoolingAssemblyDispatch.cpp new file mode 100644 index 0000000000..4a5623394f --- /dev/null +++ b/src/runtime/cpu/operators/CpuPoolingAssemblyDispatch.cpp @@ -0,0 +1,98 @@ +/* + * Copyright (c) 2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "src/runtime/cpu/operators/CpuPoolingAssemblyDispatch.h" + +#include "arm_compute/core/ITensor.h" +#include "arm_compute/runtime/NEON/NEScheduler.h" +#include "src/core/CPP/Validate.h" +#include "src/core/cpu/kernels/CpuPoolingAssemblyWrapperKernel.h" + +namespace arm_compute +{ +namespace cpu +{ +CpuPoolingAssemblyDispatch::CpuPoolingAssemblyDispatch(std::shared_ptr memory_manager) + : _memory_group(std::move(memory_manager)), + _workspace(), + _is_global_pooling_layer(false) +{ +} + +CpuPoolingAssemblyDispatch::~CpuPoolingAssemblyDispatch() = default; + +void CpuPoolingAssemblyDispatch::configure(const ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &info) +{ + const CPUInfo &ci = NEScheduler::get().cpu_info(); + const unsigned int num_threads = NEScheduler::get().num_threads(); + + // If we don't support a combination of data types, silently return: it is the caller's responsibility to check if configure() was successful via is_configured() + if(!CpuPoolingAssemblyDispatch::validate(src, dst, info)) + { + return; + } + + auto pooling_wrapper = std::make_unique(); + ARM_COMPUTE_ERROR_ON(pooling_wrapper == nullptr); + pooling_wrapper->configure(src, dst, info, ci); + + // Check if we have Global Pooling Layer + _is_global_pooling_layer = (src->dimension(2) == info.pool_size.width) && (src->dimension(1) == info.pool_size.height); + + // Allocate workspace based on kernel's memory requirements + constexpr size_t alignment = 4096; + const size_t workspace_size = pooling_wrapper->get_working_size(num_threads); + _workspace.allocator()->init(TensorInfo(TensorShape{ (workspace_size + alignment /* FIXME: remove alignment after COMPMID-1088 */) }, 1, DataType::S8), alignment); + _memory_group.manage(&_workspace); + _workspace.allocator()->allocate(); + + _kernel = std::move(pooling_wrapper); +} + +Status CpuPoolingAssemblyDispatch::validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &info) +{ + return kernels::CpuPoolingAssemblyWrapperKernel::validate(src, dst, info); +} + +bool CpuPoolingAssemblyDispatch::is_configured() const +{ + return _kernel != nullptr; +} + +void CpuPoolingAssemblyDispatch::run(ITensorPack &tensors) +{ + ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No srcs provided"); + + tensors.add_tensor(TensorType::ACL_DST_1, &_workspace); + + if(_is_global_pooling_layer) + { + NEScheduler::get().schedule_op(_kernel.get(), Window::DimX, _kernel->window(), tensors); + } + else + { + NEScheduler::get().schedule_op(_kernel.get(), Window::DimY, _kernel->window(), tensors); + } +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/runtime/cpu/operators/CpuPoolingAssemblyDispatch.h b/src/runtime/cpu/operators/CpuPoolingAssemblyDispatch.h new file mode 100644 index 0000000000..353bbe1a78 --- /dev/null +++ b/src/runtime/cpu/operators/CpuPoolingAssemblyDispatch.h @@ -0,0 +1,89 @@ +/* + * Copyright (c) 2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ARM_COMPUTE_CPU_POOLING_ASSEMBLY_DISPATCH_H +#define ARM_COMPUTE_CPU_POOLING_ASSEMBLY_DISPATCH_H + +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/IMemoryManager.h" +#include "arm_compute/runtime/MemoryGroup.h" +#include "arm_compute/runtime/Tensor.h" +#include "src/runtime/cpu/ICpuOperator.h" + +namespace arm_compute +{ +namespace cpu +{ +class ITensor; + +/** Basic function to run pooling assembly kernels */ +class CpuPoolingAssemblyDispatch : public ICpuOperator +{ +public: + /** Constructor */ + CpuPoolingAssemblyDispatch(std::shared_ptr memory_manager = nullptr); + /** Prevent instances of this class from being copied */ + CpuPoolingAssemblyDispatch(const CpuPoolingAssemblyDispatch &) = delete; + /** Default move constructor */ + CpuPoolingAssemblyDispatch(CpuPoolingAssemblyDispatch &&) = default; + /** Prevent instances of this class from being copied */ + CpuPoolingAssemblyDispatch &operator=(const CpuPoolingAssemblyDispatch &) = delete; + /** Default move assignment operator */ + CpuPoolingAssemblyDispatch &operator=(CpuPoolingAssemblyDispatch &&) = default; + /** Destructor */ + ~CpuPoolingAssemblyDispatch(); + + /** If supported create an assembly routine, else fallback to Compute Library function. + * + * @param[in] src Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. + * @param[out] dst Destination tensor info to store the result of pooling. Data types supported: same as @p src. + * @param[in] info Pooling meta-data + */ + void configure(const ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &info); + + /** Indicates whether or not this function can be used to process the given parameters. + * + * @param[in] src Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. + * @param[in] dst Destination tensor to store the result of pooling. Data types supported: same as @p src. + * @param[in] info Pooling meta-data + * + * @return a status. + */ + static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &info); + /** Was the function successfully configured ? + * + * @return True if the function is configured and ready to run + */ + bool is_configured() const; + // Run method overriden + void run(ITensorPack &tensors) override; + +private: + arm_compute::MemoryGroup _memory_group; + + arm_compute::Tensor _workspace; + bool _is_global_pooling_layer; +}; +} // namespace cpu +} // namespace arm_compute +#endif /* ARM_COMPUTE_CPU_POOLING_ASSEMBLY_DISPATCH_H */ -- cgit v1.2.1